As the prices of energy continually rise in today’s world, commercial businesses, manufacturers and home users alike are all under enormous pressure from international energy efficiency organisations to ensure their computer equipment is environmentally friendly. “Greening” equipment and operations offers companies numerous advantages not only in energy bill savings, but also in terms of reducing CO2 emissions and increasing the company’s environmental reputation (which is of inestimable value).
However, with Information Technology (IT) becoming ever more abundant within enterprises, and with a mounting need for a strong network backbone to serve and process these installations’ data, more and more electricity is required to power them. One of the more notable studies on the power consumption of office and telecommunications equipment estimated the United States’ annual power consumption
at 97TW-h in 2002 [1], an annual cost of $7.65 million† (£4.62 million). A projection of energy prices published in 2005 anticipated electricity price increases of 10% between
2005 and 2010 [3]. When this figure is coupled with the staggering adoption rate in the IT sector (an investment proportion of 40% in 1998, and rising [4]), it can be assured that energy bills will follow a similar upwards pattern.
Although several studies with enormous scope have been performed considering the “energy footprint” of office equipment and data centre operation, there has been little to no work on the subject of network-enabled devices specifically. When considering the rising importance of information and its availability in modern organisations, there is a considerable demand for the investigation of the power consumption of the devices that provide this service.
Also of interest is an examination of the literature available outlining the specification of numerous enhancements to the Ethernet standard. Being proposed at the time of writing, these technologies will result in the increased energy efficiency of existing data links. A detailed exploration of several of these breakthroughs, along other notable developments, will be contained within this report’s literature review.
Five objectives will be met by this project:
The project has been organised into the following chapters:
This chapter explores a selection of literature relating to the power consumption of network devices, first exploring existing standards that determine the power state of common device types. Also present is an examination of present and future initiatives relating to the power efficiency of devices and Ethernet data transmission.
Network devices, regardless of their type can all be considered to have certain modes of operation. Although certain types of IT equipment can only be considered “on” or “off”, more advanced devices such as workstations and network infrastructure devices can be powered down to intermediate levels where less power is consumed. Any one type of device can have different rates of energy consumption depending on the mode it is being run in.
The Advanced Configuration and Power Interface (ACPI) standard defines a set of power states for systems and devices and was developed in conjunction with major software vendors including Hewlett-Packard, Intel and Microsoft. Support also exists for migrating Linux machines to this standard [5]. Table 2.1 describes a list of system “SStates” which define the power status of an ACPI-compliant workstation.
|
ACPI Level |
Mode Name |
Function |
|
S0 |
Working |
The system is operating normally, all components are receiving power.
Although not mentioned in the ACPI standard, Roth et al. note the distinction between active-idle (where the system is not actively processing) and active-processing (where the system is performing computations [6]) and as such there can be a wide difference in power requirements from devices in this mode. |
|
S1 to S4 |
Sleep |
Levels S1 to S4 all define sleep levels of variable depth. S1 preserves most operation and conserves the least amount of power, whereas S4 provides the largest savings and powers down every possible component of the system.
Microsoft Windows computers commonly use Standby as their primary power saving mode. This mode operates at level S3. Here, user data is stored in RAM and non-essential components of the system are shut down. The CPU is provided no power, hard disks are switched off but RAM is in a constant “refresh mode” to keep the user’s data intact.
S4 mode is supported by more recent operating systems and is known as “hibernate”. This mode saves more power than its standby equivalent by saving an image of the system’s memory to hard disk before powering down, eliminating the need for the system’s RAM to be refreshed. |
|
S5 |
Soft Off |
System is completely powered down and requires a full reboot to return to S0 state. The system is still connected to the mains supply and draws a nominal amount of power (as mentioned below). |
|
“S6” (G3) |
Mechanical Off |
Not a part of the official ACPI standard, state S6 is sometimes used to refer to the global “mechanical off” state G3 [7] and implies that the supply of electricity is physically removed from the system. |
Table 2.1: S-States of the ACPI Standard [8]
It should be observed that even if a computer is considered to be in S5 mode, it will still draw a nominal amount of power. Roth’s measurements show that even when powered off (but still plugged in) personal computers and notebooks still draw 2W [6].
This phenomenon is known as “phantom load” and is common to all electronic devices. Also known as “standby power” or “vampire power”, it has been identified as a major source of energy wastage and has been a focus of many governments’ energy efficiency undertakings [9].
As such, discussions considering the benefits and drawbacks of standby modes are frequent. With the difference of power consumption between S3 and S5 modes being so small, leaving a system in standby mode overnight may be almost as energy efficient as shutting it down. Harris & Cahill go as far to suggest that power mode transitions from deeper ACPI modes typically consume extra energy (due to device start-ups) and can even reduce a system’s mechanical lifetime (due to physical wear)[10]. However, despite these discussions, it cannot be argued that putting a device into “S6 mode” garners more savings than both S3 and S5 modes. Removing a device from the electricity supply always reduces its power requirements to zero.
The ACPI model pertains mainly to personal computing systems developed by the contributing vendors. For other network infrastructure devices, the CISCO EnergyWise initiative uses a scale similar to (but greater in scope than) the ACPI model which all network devices would comply to.
2.2.2 CISCO EnergyWise
IT devices and their infrastructure are not the only consumers of electricity in an organisation. As lighting and heating alone account for 66% of an organisation’s electrical energy consumption (compared with IT equipment’s 25-30%) [11], the prospect of managing complete organisational power consumption underneath one central system is an appealing idea.
The EnergyWise initiative (developed by CISCO Systems) is a proposed energy management architecture which seeks to measure and collect power information from all its connected devices, with an aim to allow administrators to better optimize the power consumption of an organisation. It goes beyond simply conserving the power of network-enabled IT equipment and instead aspires to control all aspects of an organisation’s power usage. In order to do this, EnergyWise defines several attributes that are used to model the organisation’s system.
Similarly to the ACPI protocol discussed earlier in this review, a common language used to define power states between devices is required to standardise their management. ACPI, having applications only for PC workstations and compliant mobile devices would be an inappropriate choice for EnergyWise. Instead, CISCO developed a new set of power levels for their management system to utilise, creating a “common lexicon” [11] between devices so that power levels can be understood. In particular, this meant that existing power management standards (such as ACPI) could be mapped directly onto the EnergyWise system.
It should be noted that Table 2.2 has varying levels of complexity depending on the device it is referring to. For example simple devices such as lighting grids may only use two modes, Operational and Non-Operational. More complex devices such as PCs will have their ACPI modes married up with a “level” in the table above.
Entities represent power consuming devices connected to the EnergyWise network and can consist of several different types. Entities may be IP-based (even differentiating between Power-over-Ethernet IP and standard Ethernet IP) or not. A category exists as well for devices that operate systems unrelated to the IT infrastructure of the network, such as heating or lighting systems. Devices, no matter what type, are considered children of the EnergyWise enabled controller that they are connected to. Network switches typically act as these controllers, representing the entities that management systems will interface with in order to control the EnergyWise system.
Each entity as described above must be a member of a domain. This allows devices to be logically arranged into groups to allow more effective management of the network (in turn better facilitating its expansion). For example, sets of switches (and their children) could be grouped together based on the building floor they reside on.
EnergyWise also defines its own communication methods in order to send commands from a central management location to its devices. CISCO has suggested two methods to implement this. The first is to send messages using the Simple Network Management Protocol (SNMP) which provides a framework for network administration tasks. EnergyWise provides its own Management Information Bases (MIBs) defining how to handle data produced by the system. This allows for simple management of one switch; however Lippis notes that the limitations of SNMP make it unsuitable for managing domains containing more than one switch [11].
Alternatively, a single “Management Port” can be defined on a central controller switch that will allow administrators to gather domain-wide information by issuing commands to it. Support for requesting and changing the power levels for tens of thousands of entities is purportedly possible [11].
Attribute 5: Management Applications & API
In order to control the EnergyWise network, CISCO have provided a common API in order to allow third-party vendors to develop network management applications utilising EnergyWise information. The API allows power consumption and device efficiency data to be pulled simply from the network and be translated into meaningful colour-coded topologies. This would allow companies which have already published software controlling various aspects of a network to easily allow power-state management to the set of features offered.
The advent of EnergyWise promises to expand the role of switches within a network. Instead of switching only traditional IP traffic, switches will soon become responsible for delivering management instructions to devices or gathering reports of power consumption over a period of time. As non-IT devices are incorporated into the EnergyWise topology, switches may soon be able to perform such complex functions as alter the temperature of a building depending on the time of day. The scope for financial savings that can be gathered by a system such as this is huge, with switches being able to orchestrate the power states of devices automatically on a regular basis.
A number of initiatives have been undertaken in order to kerb the amount of energy used by IT infrastructure, network devices and more widely, electrical devices as a whole:
EnergyStar is a standard specifying power consumption requirements for a range of electronic devices. Originally created in 1992 as an American government-funded program to encourage computer manufacturers to include power management options in their products [12], it has since expanded to consider consumer and commercial products, as well as devices such as lighting and air conditioners [13]. As a voluntary accreditation, it is not required for manufacturers to subscribe to, but its high reputation amongst consumer groups provides incentive for compliance.
EnergyStar’s current fifth specification revision maintains directives on a number of different computer systems. Desktop computers, notebooks, games consoles and workstations amongst others are all included. However, server computers and more recent mobile devices (PDAs and smart phones) are not included in the specification [14]. Also of note is their specification for notebook computers which requires a lowpower mode consuming no more than 15W, which McWhinney notes that a large percentage of notebooks comply with [15].
EnergyStar has proven to be a very popular scheme, as demonstrated by its international expansion and the range of devices it now covers. In their 2006 annual report, EnergyStar reported that compliant desktop computers are shown to save between 5% and 55% more power than their non-accredited counterparts. The program in its entirety also published annual savings of $13.7 million in the year of publication, along with considerable emission reductions from the year of 2000 onwards [13]. As such, companies with large IT outlays can be assured that purchasing products accredited by EnergyStar conserves more energy and creates less carbon emissions.
Power-over-Ethernet (PoE) describes a set of standards which define a method of transmitting power over an Ethernet link whilst not disturbing the data contained on it. First published as 802.3af in 2003, a new version of the standard known as 802.3at was recently approved in September 2009 [16] featuring marked improvements to the amount of power supported devices could provide. The publication of 802.3af/at also serves to encourage standardisation of all previous work performed in the same area, such as CISCO Systems’ “inline power” technology [17].
Originally developed to provide both power and network connectivity to locations where power cabling was impractical or impossible to provide, the main advantage of PoE lies in the ability to discard the traditional AC transformer based method of supplying power to devices. PoE is of particular application to devices such as CCTV cameras and wireless repeaters (which are often positioned in out of reach locations) as well as making Voice-over-IP (VoIP) phones resemble their “plain old telephone system” counterparts more by similarly drawing power from their copper transmission line).
Two types of devices exist in the operation of PoE:
Power Sourcing Equipment (PSE): PSE equipment is typically a PoE enabled network switch which supplies electricity to connected devices. Devices known as Midspan Power Sources (MPS) are also used along with traditional Ethernet switches to “inject” power into existing Ethernet networks in the absence of a PSE switch.
Powered Device (PD): Connected devices are known as PDs, and are supplied power from the PSE via twisted pair cable. The 802.3af specification provides only around 13W of power to be supplied [18]. Whilst certainly not enough to power larger devices such as PCs and large printers, PoE has found a niche powering smaller pieces of equipment that only require nominal amounts of energy.
Despite the clear advantages that a PoE infrastructure would bring to an organisation, attention must be paid to the backwards compatibility of the platform. As thousands of varieties of standard Ethernet equipment have been deployed across the world without the ability to accept power [18], it would be foolish to arbitrarily inject power into them and risk damage or device failures. To prevent this, a discovery process is embedded into DSE devices which maintain their ports in a low-power state until devices are determined to be PoE compliant. A tentative low voltage (in the range of 2.7V to 10.1V) is then applied to a PD upon connection and the PSE checks for a built in a “signature resistance” of 25kΩ before supplying larger amounts of power [19].
Additional concerns have been raised [18] over the safety of using existing 8P8C connectors† to supply power with, particularly as the female socket is large enough for a small finger to be inserted into. However, as the 802.3af standard only provides a small DC voltage (48V) and an extremely low current (up to 300-375mA maximum) [18, 19] through the twisted pair wire, no harm can be caused.
Wake-on-LAN (WoL) is a technology designed to be used with Ethernet-compliant devices and permits them to be turned on via network communication from another device. WoL has been available for over a decade with various implementations supported by different hardware vendors [20, 21].
Prior to the introduction of WoL, computers could only be communicated with if they were in an ACPI S0 state. When technicians realised that they required a method to communicate with computers kept in other states, WoL was developed in order to “pull” a device out of its low power state and back into S0 mode.† also erroneously referred to as “RJ-45”connectors
WoL functions by requiring a device’s network adapter to remain operational whilst the rest of the device is powered down. This results in a nominal amount of “standbypower” being drawn by the device to keep it operational. The network adapter of the device would also contain software that continually listened for “magic packets”. Upon the reception of a magic packet, the network adapter would send a signal to its host, prompting it to “wake up” into S0 mode.
A “magic packet” requires a certain sequence to be contained within it in order to awaken a system. It can appear anywhere in the packet’s payload, but the sequence must take the form of six “one” bytes (each represented by FF) followed by sixteen full iterations of the device’s six byte MAC address (represented in Figure 2.1 as 11 22 33 44 55 66).
WoL could find a valuable place as part of an organisation’s power management plan. It would be possible for administrators to remotely power on machines on an as-needed basis rather than remain powered on indefinitely. However, one of WoL’s limitations is its unidirectional nature, only being able to wake systems. A worthwhile expansion of the technology would allow magic packets to shut down systems remotely. “Proxying” (discussed later in this chapter) can be considered in some regards as a more sophisticated implementation of WoL.
In October 2007, the Institute of Electrical and Electronics Engineers (IEEE) approved the 802.3az project to investigate and improve the energy efficiency of the 802.3 Ethernet standards. Its main objectives involve developing techniques for lowering the power use of Ethernet whilst retaining compatibility with the current physical media that use it.
As Ethernet is a family of technologies operating at OSI Layer 1 and 2 and is used in the majority of Local Area Networks (LANs) today, incorporating energy saving techniques into the technologies themselves will yield savings from every network that utilises them.
Currently the 802.3az project has published proposals highlighting three techniques which could increase the energy efficiency of Ethernet devices. Although distinct in their application, each proposal brings attention to the fact that most networks are kept on 24 hours a day, even when they aren’t required by users.
The first proposal published by the 802.3az project was a paper on using Adaptive Link Rate (ALR) mechanisms as a method of controlling the power usage of Ethernet links [22]. ALR was developed and refined out of the realisation that Ethernet links remain idle or in low use for a very large proportion of the time (with studies showing average Ethernet link utilisation of only 1% [23]).
Bennett proposes in his proposal that as the capacity of network media and transmitters increase so will the energy required to power and maintain the links. They observe that Ethernet links operating at 1Gbps require 2W more power at each transmitter than equivalent links operating at 100Mbps. As such they propose that during periods of low network usage, ALR would allow Ethernet links to “step down” transmission speeds in order to save power. Similarly, links would “step up” to higher rates as their services were demanded.
ALR operates from both ends of the transmission link. Both transmitter and recipient interfaces would use in-built “policies” to automatically negotiate whether data rates should be stepped up or down. Working as a handshake mechanism, a change would be made only if both parties agreed. Factors such as buffer queue thresholds and actual rate utilisation would considered in this decision.
Two scenarios would be possible:
Increase from low data rate to high: The size of the transmitter’s buffer queue is used in determining the need for a higher data rate. When over a certain amount, the burdened interface would send a frame to the recipient requesting a transition. If a higher rate is available, the request to “step-up” must never be denied by the recipient in order to guarantee maximum throughput.
Decrease from high data rate to low: The link utilisation of the interface would be monitored. If below a certain threshold, the interface would send a frame requesting a rate “step-down”. However, if the other interface’s link utilisation did not also fall beneath the threshold, the request must be denied.
Using conditions such as the above would guarantee that higher data rates would always take precedence over energy conservation. Also, as “step-up” and “step-down” requests would be implemented using a fast signalling method at the MAC level, transitions could take place promptly. This would make the amount of perceived delay negligible to the user.
As the policies for stepping up and stepping down data rates must be contained in both transmitter and receiver Ethernet controllers, both devices would have to be compliant with the ALR protocol. Unfortunately, devices in use today are not. No standard currently exists for ALR and a considerable amount of work on the “open challenges” present in the technology must be performed before one will be developed. Additionally, once a standard has been published, it must be considered whether existing Ethernet devices be able to comply with it or whether they will have to be upgraded to more recent devices. If the latter scenario is true, it will almost certainly cost a considerable amount of money for most businesses to replace every Ethernet controller present in their network. Even if it becomes possible to upgrade existing controllers, it will take a considerable amount of time until the technology is widespread enough to be enforced sufficiently to yield the massive monetary savings heralded by its authors.
However, a paper by Nedevschi et al. notes that EnergyStar standard proposals for 2009 discuss requirements for Ethernet links to use slower data rates in order to conserve energy when idle [24]. As such, ALR or a technique similar to it may see inclusion within the EnergyStar specification in the near future.
Pause Power Cycle (PPC) is a method used by LAN switches that involves adapting the power states of its own components in accordance with the states of the active links that are connected to them. The author of the technique, Francisco Blanquicet, suggests that rather than remaining powered on 24 hours a day, the main goal of switches should be to transmit data as fast as possible and then return to a “low power idle-mode” [25]. The PPC method is an implementation of this ideology.
Figure 2.2 shows how PPC might be used in a typical network. The switch periodically sends PAUSE frames to network devices and temporarily powers off the link, conserving energy. After a timer elapses, the link is then powered back on and resumes transmission of data.
Blanquicet’s initial calculations on power saving show that the energy conserved by PPC is directly related to the proportion of time it is powered down. He refers to the ratio of uptime to downtime as the switch’s “duty cycle” and cites that if it were set at a value of 50% (essentially halving its uptime), the amount of energy required by the device would be halved.
The amount of energy saved through PPC seems to depend on sacrificing network throughput. By lowering the amount of time a link is powered on, the effective transmission speed of the medium is reduced. Banquicet asserts that his technique may result in occasional buffer overflows in clients (resulting in packet loss) and his experiments with PPC’s duty cycle set to 50% show the introduction of erroneous artefacts to streamed video [25]. In high speed environments, such as LANs, this may not be such an issue, as data can be retransmitted quickly over media with large capacities. However, in wide area network environments where the available bandwidth is considerably lower, these errors suggest that PPC may need its duty cycle set to a less aggressive setting (or be disabled altogether) to provide acceptable throughput. In conclusion, this technique is a direct trade-off between link quality and device power consumption.
September 2007 saw a proposal detailing a process known as “proxying”. The concept provides a method for network terminals to be able to retain their network connectivity regardless of their power mode. An additional device (known as the “proxy”) would act as an intermediary to the terminal and preserve the network presence of its parent device.
Nordman argues that many messages destined for a workstation don’t require the use of many of its many “power hungry” components (such as CPU, hard drive and memory) and can be handled by the network interface card (NIC) itself [26]. The proxy’s main task would be to identify these messages, generate routine replies for them and determine whether the device requires to waking up. This would allow a workstation to remain in a standby power mode while the proxy dealt with maintaining its network presence.
Figure 2.3 demonstrates the process as currently proposed along with the five steps required for its operation. Three distinct entities are present: the proxy, the sleeping device and the external network.
Several different types of proxying are suggested in Nordman’s document:
Self-Proxying: Where the proxy exists as part of one of the device’s components (typically its NIC) and is controlled under the same operating system. Power would remain supplied to the proxy component whilst all of the workstation’s other components would remain off.
Switch-Proxying: Where the proxy exists as part of a network switch’s port that the device is connected to. Nordman suggests that the mobility of connected devices may pose an issue to how proxying would be implemented [26]. Existing “Wake-On-LAN” requests may be utilised in its operation.
Third Party Proxying: Where the proxy exists in a third party device such as another workstation on the network or even a dedicated device.
Proxying still appears to be in its conceptual form as an addition to the Ethernet standard. Although planned to be become a requirement of EnergyStar compliant devices [27], many of the proxying processes’ procedures have yet to be defined. In particular, its authors acknowledge that problems may arise in the implementation of switch proxying and consider third party proxying as outside the scope of their paper due to its complexity. They also concede that the fifth stage of the process still lacks the definitions for the proxy’s role after the host device has woken up.
Proxying as a concept is certainly a fascinating idea, as the infrastructure of computer networks currently have no conception of the power states of devices connected to it. The ability to place entire racks of servers into sleep mode until required would considerably reduce the amount of electricity consumed in data centres, for example. However, despite its potential to be furnished in future Ethernet hosts, its lack of maturity (and lack of a published standard) make it an unrealisable method to save power in computer networks in the near future.
Along with examining how typical host workstations represent their power states, an exploration of CISCO’s EnergyWise technology shows exactly how the future of network power management may look. As well as this, initiatives and technologies seeking to make IT equipment more efficient have been present for several years. EnergyStar has been a very successful initiative encouraging manufacturers to develop more energy efficient equipment. Wake-on-LAN technologies have also been used by network professionals for years to reduce the constant power draw of infrequently used PCs.
However, much of the future development present in network power conservation focuses on making Ethernet links themselves more efficient, fore-fronted by the IEEE 802.3az group. Despite a selection of papers exploring a range of interesting technologies, its recent inception means that standards do not yet exist for them. As such, it will likely be several years before their draft proposals are accepted by the IEEE for introduction in Ethernet devices.
Several studies [1, 12] have noted that computer systems do not draw a steady level of power over time; instead their requirements have been shown to fluctuate depending on the power mode of the system. Industry knowledge of this is reflected in the publishing of datasheets: they often contain several different power consumption figures for devices, each reflecting the level of load being put on the device.
This experiment aims to confirm that the power draw of a typical host computer does change depending on its ACPI mode, satisfying objectives 3 and 5 from one of two perspectives (the latter being explored in Chapter 4). By measuring the power drawn by a device over time, the aim is to show each mode draws progressively less power as they are powered down further. Two models of Personal Computer (PC) will be compared to show this: A machine of modest specifications, and a higher-powered workstation PC typically used for 3D rendering and video editing.
The components and software of each machine remained constant throughout the experiment with the only altered variable being its ACPI mode. The operating system of each machine was Microsoft Windows XP Professional. Datasheets for both the Optiplex GX620 and HP dc7900 Small Form Factor (included as Appendices B and C) rated the maximum output of their power supply units as 275W and 240W, respectively.
Used to measure the power draw of the system was a Maplin “Plug-In Mains Power & Energy Monitor” (hereafter referred to as the “power monitor”). It features a power measurement mode which updates every second with readings to the nearest Watt.
The following ACPI levels were used in the experiment:
The systems used had no facility to be placed into the S4 (hibernate) state, and the “S6”/G3 (mechanical off) state has been omitted due to the fact its results would always equal 0W.
The power monitor was positioned between the plug of the computer system and mains power supply and set to the power monitoring mode. The system was placed into the appropriate ACPI mode and the power readings were allowed to stabilise to ensure that transitions between ACPI modes were not still underway. Measurements were taken from the power monitor at intervals of 5 seconds for a total of 90 seconds.
|
Power Mode |
|
t |
0 |
5 |
10 |
15 |
20 |
25 |
30 |
35 |
40 |
45 |
|
|
S0 (activeprocessing) |
140W |
141W |
136W |
141W |
140W |
140W |
132W |
136W |
137W |
141W |
|||
|
S0 (activeidle) |
79W |
79W |
79W |
79W |
79W |
80W |
80W |
79W |
79W |
79W |
|||
|
S3 (standby) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
|||
|
S5 (soft-off) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
|||
|
Power Mode |
|
t |
50 |
55 |
60 |
65 |
70 |
75 |
80 |
85 |
90 |
AVG |
|
|
S0 (activeprocessing) |
136W |
136W |
132W |
137W |
139W |
139W |
139W |
135W |
140W |
137.73684W |
|||
|
S0 (activeidle) |
80W |
85W |
79W |
79W |
78W |
79W |
78W |
79W |
79W |
79.36842W |
|||
|
S3 (standby) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
|||
|
S5 (soft-off) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
|||
Table 3.1: Results for Optiplex GX620
The results of the experiment confirm what was presumed about the power requirements of ACPI levels, conclusively showing that machines in more functional states require more power. Powering the system down into S5 and S3 modes expectedly resulted in lower power consumption than leaving a machine in its S0 (Active-Idle) state.
It was also shown that the amount of power a system requires depends on how active it is: Those with more of their components actively processing can draw almost twice as much power than when idle.
The experiment also proved the presence of the “phantom load” phenomena. In S5 mode, a small amount of power (2W) was still drawn from the supply despite the fact that the device’s power button had been pressed and was presumed to be off.
Perhaps most surprisingly, the results of this experiment showed that there was no difference between S3 and S5 states. This suggests that the S3 sleep mode provided with modern operating systems is indeed a viable alternative to placing the machine in S5 state.
|
Power Mode |
t |
0 |
5 |
10 |
15 |
20 |
25 |
30 |
35 |
40 |
45 |
|
|
S0 (activeprocessing) |
56W |
55W |
54W |
54W |
54W |
56W |
53W |
52W |
54W |
54W |
||
|
S0 (active-idle) |
31W |
32W |
31W |
31W |
30W |
31W |
31W |
33W |
31W |
31W |
||
|
S3 (standby) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
||
|
S5 (soft-off) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
||
|
Power Mode |
t |
50 |
55 |
60 |
65 |
70 |
75 |
80 |
85 |
90 |
AVG |
|
|
S0 (activeprocessing) |
57W |
54W |
54W |
54W |
54W |
55W |
55W |
54W |
53W |
54.315790W |
||
|
S0 (active-idle) |
30W |
31W |
31W |
33W |
31W |
31W |
30W |
31W |
31W |
31.10526W |
||
|
S3 (standby) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
||
|
S5 (soft-off) |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
2W |
||
Table 3.2: Results for HP dc7900 Small Form Factor
3.3.4 Discussion
The results from a lower specification PC confirm the conclusions reached above. Power requirement increases are, however, of a lower proportion than the higher specification machine due to less power being consumed overall.
Phantom load is again confirmed with the S5 reading remaining at the same amount as before (2W). This suggests that phantom load is either constant, being unrelated to the capabilities of the system, or that it is more related to the nameplate value of the system’s PSU (both of which are very similar between these systems).
Again, there was no measured difference between S3 (sleep) and S5 (soft-off) modes, confirming that S3 can be used as a viable alternative to turning machines off.
Between the two machines examined, it was clear that machines developed for different applications have dissimilar power requirements. Unsurprisingly, machines geared towards heavy computational tasks (such as rendering 3D models and editing video files) required more power when placed under strain. The reason for this is most likely due to their more powerful (and power-hungry) processors and graphics chipsets.
Nameplate values on the power supply seemed to have no relevance to the amount of power that a machine actually used, even when placed under heavy load. As such, nameplate values should not be used in power consumption calculations in an organisation, as the power consumption would be grossly overestimated. Instead, readings from a meter should be used.
The results also show that S3 and S5 ACPI levels offer large power saving opportunities for network devices. Enterprises stand to conserve considerable amounts of energy by placing devices into S3 or S5 whenever their use isn’t required (anywhere thirty to seventy times, depending on the machine). S3 mode would be recommended for this, as it allows faster resumption of service with no increased power overhead. For large organisations, these transitions needn’t be performed manually: technologies which could place devices into these states remotely (such as CISCO EnergyWise) could aid IT Staff in automating these transitions, helping to provide maximum savings.
In order to allow communications between large numbers of IT devices, the use of numerous hubs, switches and routers are required. However as noted by Coffman & Odlyzko [23], Ethernet links find themselves sitting idle most of the time. The purpose of this experiment is to determine how much power these idle links use in comparison to those which are under constant stress, fulfilling Objectives 3 and 5 from an alternative perspective.
Twelve Optiplex 755 computer systems (running a virtualised Linux operating system) were connected to the twelve 10BASE-T ports on a CISCO Catalyst 1900 LAN switch with CAT5 cable. This particular model can be considered typical of a low-end access layer switch, often widely deployed to enable connectivity to hosts devices on a network. The default configuration of the switch was used, and connectivity between the hosts was confirmed using the “ping” command. A data sheet for the switch is included as Appendix D.
All connected devices were considered peers with no distinction between “server” and “host”. This configuration was chosen to mimic the operation of the access switching layer of a typical network.
The variable factor in the experiment would be the “utilisation” of the router. The CISCO Catalyst 1900 switch contained a control on its front face that allowed access to a UTL mode that displayed how much of the switch’s bandwidth was being consumed on the light emitting diodes (LEDs) above its ports. Table 4.1 shows how the switch represents load on its LEDs.
Three different power states were defined:
Idle-Disconnect: All CAT5 cabling was disconnected from the switch in order to gain a baseline power consumption reading for the switch. In this mode, the connected devices had no connectivity to the switch or each other.
Idle: The twelve hosts were connected to the switch while performing no user-initiated communications tasks with either the switch or each other. Operating system “housekeeping” tasks, such as the detection of an active connection, would be performed but their low overhead meant that any load put on the network could be considered negligible. The UTL mode of the switch should display a maximum of one illuminated LED.
Under Load: Each of the devices connected to the switch were instructed to send large amounts of data as fast as possible another device on the network using the following
command:
ping <ip-address> -fs 65507
<ip-address> denotes the IP address of the destination host.
-s 65507 denotes that packets should be the maximum ICMP packet size of 65507 bytes, plus 8 bytes of ICMP header data (resulting in a total packet of 65515 bytes).
-f denotes “flood ping” where ICMP ECHO_REQUEST packets are sent as fast as responses are returned (or at a rate of one hundred packets per second, whichever is more) [33].
Using this scheme, it can be shown that each port on the switch would be working at several times its maximum capacity:
UTL = P × b
UTL = 100 packets/sec × (65515 × 8) bits/packet
UTL = 49.98 Mbps
Where:
UTL = Utilisation one of the switch’s ports in Mbps
P = The number of packets sent per second (minimum value of 100)
b = The number of bits per packet
Since the maximum data rate of each port was 10Mbps, it should be expected that large amounts of packets would be dropped across the switch. This was confirmed by checking the status mode (STAT) of the switch, which showed alternating green and amber lights.
Using the above scheme it was possible to increase the UTL mode of the switch to ten
LEDs. With reference to Table 4.1, this meant that more than 20Mbps and less than
140Mbps of the switch’s bandwidth were being used.
First, the power monitor is placed between the plug of the switch and the mains power and set to the Watts monitoring mode. The router is then placed in its desired “power state”. For the “Under Load” state, thirty seconds are allowed to elapse to ensure that network traffic is being transmitted across the switch. Measurements were taken from the power monitor at intervals of 5 seconds for a total of one minute.
|
Load State |
t |
0 |
5 |
10 |
15 |
20 |
25 |
30 |
35 |
40 |
45 |
50 |
55 |
60 |
|
Idle- Disconnect |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
14W |
|
|
Idle |
15W |
15W |
14W |
15W |
15W |
15W |
15W |
14W |
14W |
15W |
15W |
15W |
15W |
|
|
Under Load |
16W |
16W |
16W |
15W |
16W |
16W |
15W |
15W |
16W |
15W |
16W |
15W |
16W |
|
Table 4.2 Results of Experiment
The results of the experiment show conclusively that the switch draws more power from the mains supply whilst under significant load. However, the difference in power draw between the under-load and idle modes is very small, having a maximum difference of two watts and occasionally even drawing the same amount. When completely unplugged from all hosts, the switch consistently drew a lower amount of power, implying that there is a minimum amount of power required simply to keep the switch powered on. This means that the overhead required in order to transmit data across a switch is in the range of one to two Watts.
In the context of a large organisation, one or two Watts per network infrastructure device is barely discernible in comparison to the power saving potential of other device types, such as PCs and display devices. This means that although technologies such as Active Link Rate and Pause Power Cycle can offer power saving potential for network devices, more effective results could be produced elsewhere.
In order to better understand how power is used in existing network devices and determine what impact any proposed energy-saving schemes would have on them, it is necessary to appreciate the fundamental power requirements of digital data transmission. By calculating exactly how much power is required to transmit data it can be seen whether the transmission of data itself is the main consumer of power in a network device, or see whether the overheads required to operate other components of the network device itself are the main consumers. This chapter fulfils the requirements of Objectives 2 and 4, first investigating theoretical power requirements and secondly comparing them to results already achieved.
This section will outline the low-level operation of the 10BASE-T technology, which operates at 10Mbps. Although 10BASE-T may not be considered a “modern” technology in terms of its throughput and sophistication (especially with the succeeding 100BASE-T and Gigabit Ethernet technologies becoming increasingly available), it has been chosen to model the power consumption of data transmission for several reasons:
This section, although not representative of all Ethernet technologies currently available, can be considered to provide a basis for further work in this area. The calculations performed were made based on thorough research of the Ethernet 802.3 standard and its related documents. They are entirely theoretical in nature, and although they do successfully represent results that were recreated in an experimental environment, it would be advisable to verify the above model before adapting it for any more elaborate technologies.
Figure 5.1 shows an arrangement of the subsystems of the 10BASE-T technology at layer one of the OSI model (also known as the PHY). Of particular interest are the items incorporated within the Medium Attachment Unit (MAU). The MAU represents a
collection of further subsystems that are central to the transmission of data over a particular medium and can be considered analogous to a “transceiver”. In this way, the MAU and its subsystems are directly responsible for the encoding of data passed from the Attachment Unit Interface (AUI) into low-level electrical impulses for transmission across the attached medium.
The MAU itself is further split into two subsystems, the Physical Medium Attachment (PMA) and the Medium Dependant Interface (MDI).
The PMA acts as an intermediary between the higher level Physical Signalling PLS system and the medium itself, translating the messages received into a form fit for transport over the MDI.
The MDI defines exactly how the media used between two MAUs is connected, detailing items such as male and female connector types, wiring diagrams for these connectors and specifying transmission and receiving sections of the medium.
As far as 10BASE-T is concerned, the following items are specified:
Connectors: A MAU MDI “connect” interface is defined, specified to accommodate a connector which resembles (but does not name explicitly) the 8P8C connector which has become intimately associated with the Ethernet technology.
Wiring & Media Use: Although specified earlier in the standard, it is made explicit here that 10BASE-T is meant for operation with copper twisted-pair links. Each medium should contain four pairs of wires, although only two of these pairs are required for this technology. Two different wiring schemes are specified for the media termination connectors: One for standard communications, and one “crossover” variety for connection of devices which operate on the same OSI layer. Individual twisted-pairs are also identified for particular functions, one for transmission and one for reception:
In 10BASE-T, the operation of data transmission can be considered as a series of AC voltages (in Manchester encoded form) supplied by a transmitting MAU to be transmitted down an Ethernet segment. This is considered to take place over one twisted pair of a CAT5 medium (typical of 10BASE-T links), with the remaining active pair being used for signal reception (as defined in the MDI). The Manchester encoded symbols are taken to have particular voltage values, and each wire that comprises part of a twisted pair is considered to have its own distinct resistance (referred to as the characteristic impedance).
10BASE-T Ethernet uses Manchester code as its data encoding scheme. One of the main characteristics of Manchester encoding is its ability to embed a clock-signal within its data stream, helping to ensure that both the transmitter and recipient of the data remain synchronised with each other with a high degree of accuracy. Unlike other encoding schemes, Manchester code represents each bit of data by a transition, meaning that two voltage levels are required to represent one bit.
Although offering added reliability and error detection capabilities to the transmission when compared to other data encoding schemes, this means that Manchester encoding requires twice many symbols to represent the same amount of binary encoded data (essentially doubling the bandwidth requirements of the data).
In 10BASE-T Ethernet, the two voltage levels used to represent Manchester encoded symbols are, on average, 2.5V and -2.5V [36]. As the transition is almost instantaneous (with the delay between being taken as negligible), a voltage (either positive or negative) is being transmitted whenever data is being sent.
Although the bottom signal does have considerably more transitions due to the need to “reset” the cycle to -2.5V at the start of each “1” symbol, it is possible to see that a constant voltage is being transmitted no matter what the content of the message (due to 0V never being used). As such, these sample signals show that a voltage of 2.5V is continually applied to the circuit shown in Figure 5.3.
Confusion exists about the voltage levels used by 10BASE-T’s Manchester coding. Several sources that speak about 10BASE-T Ethernet attempt to define the voltage values for Manchester encoded signals as 0.85V and -0.85V respectively (including, notably, Andrew Tannenbaum [38]). However, the 802.3 standard makes no reference to this figure at all. As such the mean values of 2.5V and -2.5V as shown in the 802.3 standard are taken as representative of a realistic installation.
As an AC circuit simply consisting of two end points, the impedance of the twisted pair’s individual wires is needed to calculate power use. Impedance is a measure of opposition to AC signals and can be considered similar to “resistance” in traditional DC circuits. Impedance is frequency dependant and the frequencies used vary depending on the technology. In this case, 10BASE-T Ethernet uses frequencies between 5MHz and 20MHz [39].
Suggested CAT5 cable specifications recommend the following impedance values in a finished product:
“Finished cable shall have a characteristic impedance of 100 ohms ±15% in the frequency range from 1 MHz to 155 MHz when measured in accordance with ASTM D 4566 Method 3” [40]
[referring to a cable length of up to 100m] Consequently for the following calculations, the impedance of the medium is taken as the range of 85Ω to 115Ω at the frequency of 10MHz.
The definition of impedance implies that it is a complex number consisting of both a magnitude and a phase [41]. However, according the research conducted, the impedance of CAT5 cabling has been referred to as one particular number (assumedly its magnitude). As such, the above calculations do not take into account the relative “phase” of CAT5’s impedance and any effect it may have on power consumption.
Now that all the necessary pieces of information have been gathered from the circuit in question, it is possible to find an appropriate equation to calculate the power usage of the circuit. In order to calculate the power requirements of 10BASE-T data transmission, the following equation is considered, where P equals Power in Watts, V equals Voltage in Volts and R equals Resistance in Ohms (Ω).
It has been mentioned that the characteristic impedance can be considered similar to resistance, in a traditional sense. In fact, it has been shown that it is possible to substitute this value into traditional power equations, such as the above example, considering it analogous to resistance [41]. And so, by substituting this value into the above equation, it changes to the following (with the symbol Z representing characteristic impedance):
The Manchester Encoding examples in Figure 5.5 show that no matter what data bits are transmitted, a constant voltage of 2.5V is being applied to the circuit. This value, along with the range of possible impedances at the designated frequency (values of 85Ω to 115Ω) was used to calculate the amount of power required to constantly transmit data.
For the range of impedances specified for CAT5 cabling, power values from 54.348mW to 73.529mW were obtained. These figures represent the power cost of sending data from one station to another in half-duplex mode. Realistically, stations utilising 10BASE-T will be operating in full duplex mode, doubling the power requirements. As such, it would be pragmatic to take the power requirements for one port to be in the range of 108.696mW to 147.58mW.
On a twelve-port switch, such as the CISCO Catalyst 1900 examined in Chapter 4, the total additional power requirements to facilitate full duplex transmission over all twelve ports will be in the range of 1.304W to 1.770W. This figure represents just the amount of power required to enable data transmission and does not take into account any of the other overheads required by the switch.
The work performed above is only pertinent to 10BASE-T type Ethernet. Although still commonly deployed across computer networks, it is important to realise that more recent, more elaborate technologies are currently leading the market.
Fast Ethernet (100BASE-TX), for example, operates at 100Mbps and could be described as the de facto Ethernet standard in operation today. Unlike 10BASE-T, it utilises a different encoding scheme (4B/5B), meaning that the voltage levels used to represent data bits would be different. It is also expected that as more data is being transmitted by this technology that power requirements would rise.
Further to this, Gigabit Ethernet (1000BASE-T) is likely to become increasingly popular as computer systems yearn for more and more bandwidth. Again, its encoding scheme differs from its predecessors, using PAM-5 encoding to allow the transmission of four symbols in parallel [38]. This increased throughput would be likely to be reflected in the power usage of the technology, increasing it significantly.
Some articles [42] refer to a non-standard 1W per port target that Gigabit Ethernet ports aspire to achieve, around ten times the power requirements of a 10BASE-T port. From this it can be seen that as the data rate of a technology increases, so will its power requirements.
Also of note is the emergence of Power-over-Ethernet switches. As the ports from these devices are expected to provide a considerable amount of power via the connected CAT5 cabling (around 13W via 802.3af), the power requirements of devices that offer this capability will be significantly higher.
As such, the results calculated for 10BASE-T should not be taken as entirely conclusive or as particularly up to date, as a plethora of other widely used technologies may prove to have wildly different requirements. These further technologies could make for interesting further study on this subject, however.
From the results obtained from these calculations, it can be seen that the task of data transmission using 10BASE-T does not actually require much power at all. Only one or two Watts would be required to both transmit and receive data on a twelve-port switch, an immaterial amount of power compared to the overheads required to actually operate the switch.
Although power savings at this low level may be relevant to, for example, portable battery powered devices (where every Watt matters when attempting to increase battery lives) the savings garnered by forcing mains-powered devices to be more efficient at the data encoding level will prove miniscule. Consequently technologies such as Pause Power Cycle (discussed in Chapter 2), although novel and effective at reducing power by small amounts, may not be worth the detrimental effects they introduce to network traffic.
The results obtained from these calculations are consistent with those observed and recorded in Chapter 4. Despite as much load as possible being put on the 12-port network switch, the power reading was only observed as increasing by one watt at the most, making the results of these calculations viable. If it were possible to completely consume all available bandwidth, it would be expected to see this power consistently lie around 2 Watts higher than its idle-disconnect mode.
Due to the resolution of the readings provided by the power monitor used (which took power readings to the nearest full Watt), it was not possible to compare the jumps in power consumption in any great detail: for future iterations of Chapter 4’s experiment, it would be desirable to locate a power monitor that could measure power changes down to the mW.
This chapter outlines a case study of power consumption throughout a typical organisation with a medium-to-large IT installation. A thorough investigation of many device types was undertaken in order to gain a broad picture of how power is used (and in what amounts) across the organisation, satisfying Objective 1 of the project.
The first step towards conducting a study on typical organisational power consumption was to choose an appropriate organisation. Several requirements were outlined:
It was decided early in the project that the University of the West of Scotland would satisfy the first requirement: a University campus requires constant connectivity for both staff and students internally, as well as upholding a public presence on the Internet. It was clear that the IT infrastructure of the campus was an elaborate and intricate installation. The presence of the School of Computing in the University added to the conception that there would be a wide and varied set of devices deployed across its network.
The University itself consists of four separate campuses: one each in Paisley, Hamilton, Ayr and Dumfries. These campuses’ IT installations are all interconnected, allowing communication over a vast geographical distance. Each campus is also connected to an off-campus data centre where large amounts of data are stored and served.
In order to satisfy the second requirement, the scope of the study was reduced to the Paisley campus alone. It was clear that an extensive study into the power requirements of the inter-campus network would be out with the time constraints of this project.
Requirement 3 was able to be satisfied as well: student labs and classrooms provided ample equipment for readings to be taken from, and assistance from the IT Services team at the University provided information about some of the more opaque aspects of the network that would have been otherwise inaccessible.
In order to complete the study within the time constraints of the project, it was deemed impractical to create a unique methodology to model the power consumption of the campus. To satisfy both time constraints and provide the most accurate portrayal of the University’s network, an existing methodology developed for measuring organisational power consumption would be used. Thus, before any practical readings were taken in the campus, an analysis of existing power consumption studies was performed.
The main requirements of a suitable study would be to first of all be relevant to the University’s range of devices, and second of all be resilient to adaptation by this project. An additional concern would be the resilience to estimation, should completely accurate figures and measurements be unavailable.
Upon initial research, it seemed there was little work dedicated to analyse how much power network devices use. Rather, many power consumption studies were carried out with much wider scope, generally focusing on the power consumption of an entire country, or in smaller cases, single data centre installations. A selection of relevant literature is evaluated below.
The 2002 publication on the energy consumption of office and telecommunications equipment by Roth et al. (hereafter referred to as Roth, for simplicity) consists of a 211 page study which examines the Annual Electricity Consumption (AEC) of a range of office equipment categories. Its broad scope encompasses many items of interest to this project including personal computers, server computers, display devices, printers and computer network equipment but also discusses the impact of other devices such as uninterruptable power supplies (UPS), copiers, telephone networks, et cetera.
Its analysis of office equipment’s AEC ratings found that network infrastructure equipment used a comparatively small amount of power compared to other office equipment. Chart 6.1 shows that computer networks and their associated devices only use 6.4TW-h (terawatt hours) of the total 97TW-h consumed by office equipment in the year 2000 (6.6%).
However, Roth did identify network infrastructure devices as an area worthy of further investigation, devoting an entire section of their study to their impact. Notably, Roth subdivides the area of network infrastructure equipment into distinct device types, considering hubs, switches (both LAN and WAN) and routers as separate areas.
Traditional LAN equipment was shown to use the majority of power in this area with hubs, LAN switches and routers claiming 6TW-h (6.4%) of power consumption figures.
More specialised equipment such as Cable Modem Termination Systems, Remote Access Servers and WAN switches were also shown to have a comparatively small (3.6%) combined contribution to power consumption.
Chart 6.1 also shows the impact that PC and server computers have on energy consumption (over 30% combined). As these devices are can be considered as endpoints on a network, their power consumption is also relevant to this project. Likewise, printers are often connected to modern networks and as such their 5.9% contribution can be considered as worthy of investigation.
Below, a sample of Roth’s study is examined in order to gauge its appropriateness for this project. The section chosen details Roth’s calculations of the AEC of network infrastructure devices.
6 . 2 . 1 . 2 S A M P L E A N A L Y S I S : N E T W O R K D E V I C E S
Network Hubs
AECHUB = N× Pport × tOH
Network hubs were shown to use 25% of the total network infrastructure device AEC figure. Roth’s methodology for measuring the AEC of hubs consisted of calculating a watt per port value for each device (Pport). An average power draw value for the entire device would be taken and then be divided by the number of ports present on the device. The resulting value would allow larger capacity devices which used more power to be compared fairly to smaller, less energy consuming ones.
† Figures are subject to rounding
Despite this attempt to levelling the playing field, Roth’s findings showed that larger capacity devices used a lower amount of power per port, with an 84-port hub using 1.23W/port, while a 96-port hub used only 1.13W/port. Both hubs were from the same manufacturer.
In order to calculate the AEC of the hub, a generous value for Pport was used (1.25W) to account for the variety of hub models deployed across the country. This value was then multiplied by industry estimates for the number of ports installed in all commercial buildings in the United States (N) and multiplied by the number of hours in operation per year (tOH). In his calculations, Roth realised the necessity of computer networks being available at all times. As such, his operational hours are always taken to be the “always on” value of 8,760 hours per year. The resulting value of these calculations could be considered the AEC value of the all hubs in the country.
LAN Switches were shown to use 52% of the total network infrastructure device AEC figure. A similar method was used to calculate the AEC of these devices, with Roth’s findings showing that switches tend to use more power per port than hubs, with an average Pport value being taken as 4W.
AEC of Routers
AECROUTER = N× PAV × tOH
Routers were shown to use 17% of the total network infrastructure device AEC figure. Roth used a different methodology for calculating the AEC of these devices. As routers do not generally have as many ports installed as switches and routers, a power per port value would be misrepresentative. Instead, he simply considered an average power draw for a typical router (PAV, taken as 40W) and multiplied that by an estimated number of routers in operation (N) and the same constant tOH value of 8,760 hours.
AEC of WAN Switches, Cable Modem Termination Systems & Remote Access Servers
AECWAN = N× P × tOH
These devices collectively represented 3.6% of the total network infrastructure device AEC figure. The first of these, WAN Switches are used to manage WAN traffic, with a typical application being the aggregation of DSLAM (itself multiplexed DSL traffic [1] traffic for ISPs. Roth touched only briefly on the methodology used to calculate the AEC of WAN switches. He abstracts the stock of devices into a number of “shelves”, representing the typical method of distribution among vendors. A simple power measurement is taken as representative of all devices here. Again, the concept of always-on computing is represented by tOH being taken as 8,760. The method of calculating Cable Modem Termination Systems and Remote Access Servers was not discussed in his report.
Roth’s analysis of network devices identifies LAN switches as the largest consumers of electricity in the network infrastructure device area. Since the number of hub ports and switch ports is very similar (93.5 million hub ports [28] to a mean of 92,500,000 switch ports †) , the main reason for this is the fact that Roth’s investigations showed that switch ports tend to use more than three times as much power as hub ports.
One disadvantage of Roth’s report was the amount of estimation required in gathering an inventory of each type of device. Because the scope of the study was so huge (calculating AEC values for devices deployed across all of the United States), the margin of error in estimating amounts of devices would no doubt be considerable.
The sources Roth cited in his estimations also tended to be published comparatively far apart. His estimations for hub ports were based on a report carried out by Silva in 1998 whilst his switch port estimates were gathered over 1999 and 2000. It would be expected that a lot more hub ports would be installed over 1999 and 2000, something his AEC calculations should reflect. As a result, Chart 6.2 should show an increased proportion of power being consumed by hubs.
Of additional concern, the power per port values calculated for all of these devices would be gathered from only one or two different models of device. This abstraction fails to represent the diversity of devices deployed across the country. As such, his power per port values could be misrepresentative of the country’s actual average.
† Studies showed a range of 90,000,000 [29] to 95,000,000 (ADL Estimate based on [29]) switch ports in operation in 1999/2000
However, Roth’s methodology would be extremely accurate if used in a context where exact inventory, power draw and model types of network devices were known. This study also contains useful measurements for considering the contribution of workstation PCs and servers and would allow a diverse analysis of the University’s power usage to be made.
In 2005, Sun & Lee examined in detail the power consumption of two data centres and found them to be facilities that consumed large amounts of energy [30]. Interestingly, they noted that the energy requirements of data centre floor space (per m2) could exceed that of traditional commercial office space by eighteen times †.
Sun & Lee’s study differed from Roth’s considerably, most notably in that the devices examined were abstracted considerably more. Also present was a more detailed examination of lighting circuits and heating, ventilation and air conditioning (HVAC) systems which were only briefly touched on in Roth’s study. They considered the data centre to have four broad areas of power consumption:
† Commercial office space typically measured 50W/m2 to 110/m2 while data centre power demand had a much wider range of 120W/m2 to 940W/m2 [30].
|
Area |
Description |
|
IT Equipment |
Defined by Sun & Lee as “servers, data storage, network devices, monitors, etc” [30], the devices which provide services to users. |
|
UPS Loss |
The wastage of power directed into the Uninterruptable Power Supply. Whilst giving power redundancy to all devices they are connected to, UPS efficiency varies greatly. |
|
HVAC |
As IT Equipment generates a lot of heat in its operation, facilities are required to regulate the temperature of a room and ensure its proper ventilation. |
|
Lighting |
The energy impact of overhead lighting used by staff in the data centres. |
Table 6.1: Sun & Lee’s Device Criteria
This study also asserts the conception of “always on” computing, with both data centres showing that their IT equipment (along with supporting HVAC and UPS devices) were kept powered on 24 hours a day, seven days a week [30]. Only lighting equipment was powered on or off on a scheduled basis, dictated by its occupancy.
The major findings uncovered from Sun & Lee’s study was that supporting the operation of IT Equipment often consumed more energy than the IT Equipment itself. Chart 6.3 below shows a breakdown of the entire energy consumption of a data centre that they examined. From this we can see that just over a quarter of data centre power usage supplies the devices themselves, and that the remainder of energy is used in providing stable operating conditions for devices (HVAC), visibility for users (Lighting) and redundancy of the power supply in case of failure (UPS).
Sun & Lee made several recommendations on how to reduce energy expenditure in data centres. Most of these suggested the reconfiguration of the support services to be more efficient. They reinforced the necessity of keeping the ratio of support services cost to IT Equipment cost as low as possible by saying “Generally, a larger contribution from the IT equipment to the total energy use indicates a better overall energy performance” [30]. No recommendations were made on limiting the operation of IT equipment (such as turning off or suspending workstations when not in use).
Several criticisms can be made of Sun & Lee’s study which could question the reliability of their conclusions. The first item of note is that both data centres were located in Singapore. Due to the tropical nature of the temperatures in this country, coupled with the fact that their study was carried out in the middle of summer, it could be suggested that HVAC systems would be much more abundant in this country (and more under
load at this time) to keep the temperature of the data centres at an operable level. Consequently, a similar data centre located in a more temperate region could see the HVAC contribution decreased.
Sun & Lee’s methodology in measuring power consumption between the two data centres could also be seen as slightly misleading. The metrics for graphing both centres’ energy use failed to take the floor space of the facility into consideration. Considering that data centre 1’s total floor space was 97m2 and data centre 2’s floor plan was more than ten times that at 1048m2, different HVAC and UPS requirements for larger premises might explain the seeming “inefficiency” of data centre 2.
Sun & Lee also make no differentiation between network infrastructure devices and workstation computers. Indeed judging by their definition, even printers, monitors, projectors and scanners could be included in the final energy measurements. This meant that no real conclusions regarding the power consumption of network devices could be made.
It would be interesting to see Sun & Lee’s methodology put to use in a larger range of data centres. In this report, only two facilities were investigated. However if the trend of IT Equipment’s energy consumption being over shadowed by support services was as consistently high as shown in Chart 6.3, then seeking to increase HVAC and UPS efficiency would perhaps be a more worthy task.
Roth’s methodology for the AEC calculations of Office and Telecommunications Equipment has been identified as the most appropriate choice. It offers the flexibility to analyse a range of devices present on a network and includes a detailed examination of network infrastructure devices. The methodology, although broad in scope, offers the potential to be scaled down for the purposes of this project. With its heavy bent towards estimation, it would also be forgiving if complete data for the University could not be obtained.
A number of device types have been omitted from the following study, originating from Roth’s report, the report examined by Sun & Lee, or items observed on the University’s network.
The reasons for omission vary depending on the area, but largely are a combination of irrelevance to the area of network power consumption and the time constraints imposed by the project.
While none of these devices would take up a particularly significant segment of any results gained, in more detailed future study it may be advisable to implement them. However, it must be mentioned that each of the areas below will impact on the power consumption figures of the organisation studied: Investigations made in the future focusing on the energy conservation of office equipment or the efficiency of data centres should take note of these areas and implement them.
Supercomputers
Knowledge of or access to a supercomputer in the University could not be gained.
Magnetic Disk Storage Systems
Knowledge of or access to these systems in the University could not be gained.
Copy Machines
Copy machines cannot be truly considered as part of the University’s computer network, often being stand-alone devices for independent use. Also, their use in the University is perhaps not as ubiquitous as in the offices that Roth’s report is based on.
Uninterruptable Power Supplies
Although most certainly deployed in the University (most likely in combination with server computers, to provide power redundancy in event of a disaster), they are not in them selves network devices. They take up only a small portion of Roth’s report, accounting for only 5.8% of office equipment energy usage.
Telephone network equipment
Since the University does not yet utilise VoIP for its telephony network, this section is irrelevant to the University’s computer network deployment.
Although essential in supporting the network’s infrastructure devices and undoubtedly a factor in the financial upkeep of these services, Roth’s study treats HVAC systems as out of scope and of “considerable complexity”, and for the same reason this study will do the same. Sun & Lee’s study does examine these systems in considerable complexity, so reference to their study is recommended if knowledge on this area is required.
Although mentioned in Sun & Lee’s methodology and considered a major contributor to the power footprint of a data centre, the investigation of the cost of lighting was not deemed relevant to the objectives of this project.
Wireless connectivity in the University, although accessible at several points throughout the University, cannot be considered as fully functional. Also, the locations of access points and the deployed number of these devices would be fully unknown, as most are hidden out of plain sight.
Intrusion Prevention Systems (IPS) / Intrusion Detection Systems (IDS) These devices are not acknowledged in Roth’s report, nor were they available to take readings from. Although they do impact slightly on the network’s power usage, they are considered outside the scope of this report.
Similarly to IPS and IDS devices, these are not acknowledged in Roth’s report, nor are they readily accessible to take readings from. Although they do impact slightly on the network’s power usage, they have been excluded from this report for simplicity’s sake.
Using Roth’s methodology, the resultant study estimates that the Paisley campus of the University of the West of Scotland consumes approximately 1229MW-h of electricity per year in order to keep the University’s network operational. It is worth noting that if the out-of-scope areas noted in Section 6.4 were taken into account, the actual power usage of the campus would be slightly higher.
The most significant contributors to the University’s power expenditure are undoubtedly the PCs and workstations that necessitate the network’s existence, comprising over a third of energy consumption (485.35 MW-h). This is not surprising, as there are in excess of two thousand machines connected to the network, a large proportion of which remain powered on continuously.
In order to provide around-the-clock connectivity to the many services offered by the network, the underlying network infrastructure equipment must also be continually powered. These devices are not limited to those access layer switches connected to enduser workstations, but also take into account core switching and routing functions. This requirement to provide constant connectivity is represented by a considerable power outlay: network infrastructure devices utilise approximately another third of the total power consumption (418.49 MW-h).
The remaining third of power used in the University’s network is split between three further areas: the monitors and displays which accompany every PC (150.50 MW-h), printers connected to the network (97.99 MW-h) and server machines which provide the network’s internal and external services (76.91 MW-h).
The remainder of this chapter is split into sections each representing a sector of Chart 6.4. Within each section is a description of the devices examined, how the stock of each device type was obtained, how the typical operating hours for each piece of equipment were calculated and the typical power requirements of each device type. Finally, calculations detailing the AEC of each device type are made, along with any relevant accompanying observations. Any recommendations that could curtail the consumption of energy on the campus are appended to these observations. Concluding this section is a comparison of the results of this study with those obtained by Roth in his report.
6 . 5 . 2 . 1 B A C K G R O U N D
The desktop computer is one of the most ubiquitous elements of an organisation’s network. Indeed, the computer networks were originally developed to facilitate the sharing of information between computers. As the personal computer can be considered as a network device itself (referred to as “building blocks” [31] of a network), an investigation of the power consumed by typical machines is of interest when gauging an entire organisation’s network’s power usage.
6 . 5 . 2 . 2 A E C C A L C U L A T I O N F O R P C S
Stocks of machines distributed across the campus were obtained from a representative of IT Services (included as Appendix E). These figures include computers contained within student accessible laboratories as well as staff computers connected to the University network. It should be stressed that only computers connected to the network could be accounted for and that an unknown amount of non-networked workstations were likely to exist. However, as this study concerns itself only with the power consumption figures for the campus’s network, the numbers obtained were adequate.
An exchange with IT Services disclosed that there was no power conservation plan in place for student lab workstations on the campus’s network. Individual power schemes of machines did not include instructions to place machines into lower power schemes after a certain amount of time had elapsed. Likewise there were no plans to power down PCs over holidays and other periods of no use. As such, the times for “standby” and “off” modes were taken as 0 hours per year.
Speaking with security staff of the University revealed that student lab machines are never powered down, even when the labs are closed. Taking this into account, student machines were assumed to be powered on and in active mode for 8760 hours per year.
Staff machines were assumed to be on for as long as staff members were in the University, and powered off otherwise. Since a wide range of working hours exist for the various faculties, 2,610 hours annually was taken as a generous estimation for each machine (ten hours per weekday).1
|
Type of Machine |
Active (S0) |
Standby (S3) † |
Off (S5) |
Unplugged (“S6”) |
|
Student |
8760 |
0 |
0 |
0 |
|
Staff |
2610 †† |
0 |
6150 |
0 |
Table 6.2: Usage times per power mode, per PC (hours/year)
† Roth’s report made distinctions between standby and suspend modes. Since the computers in the campus use neither of these modes for any length of time, they have been combined as “Standby (S3)”. †† 261 weekdays in 2010
For student lab machines, two types of PC were considered, standard specification lab machines and the higher end workstations contained in high performance labs. An average for the power consumption of each of these machines was used to one student machine. Staff machines were taken as being of standard lab machine specification. The power draw information below has been reapportioned from Chapter 3.
|
Type of Machine |
Active (S0) |
Standby (S3) |
Off (S5) |
Unplugged (“S6”) |
|
High Spec Lab (Optiplex GX620) |
79.37W |
2W |
2W |
0W |
|
Standard Spec Lab (HP dc7900 SFF) |
31.11W |
2W |
2W |
0W |
|
Average |
55.24W |
2W |
2W |
0W |
Table 6.3: Power draw of typical PCs in each power mode
|
Type of Machine |
Installed Base |
Mode |
Use (h/year) |
Draw (W) |
Annual AEC (MW-h) |
|
Student |
684 |
Active (S0) |
8760 |
55.24 |
330.9892416 |
|
Off (S5) |
0 |
2W |
|||
|
Staff |
1,651 |
Active-Processing (S0) |
2610 |
31.11 |
154.3637121 |
|
Off (S5) |
6150 |
2 |
|||
|
|
|
|
|
Total |
485.3529537 |
Table 6.4: AEC Calculation for PCs
PCs and workstations in the University are estimated to use around 485MW-h per year.
The most evident observation that can be made by the figures gathered are that despite being almost three times fewer in number, student PCs located throughout the University consume more than twice as much power. One of the reasons for this is that a portion of student machines themselves are higher consumers of power than staff machines with their more humble specifications. However, the main reason for the sheer amount of power being consumed is due to their hours of operation; being kept in active mode throughout the year (even throughout the night!) results in a vast increase in electricity usage.
A recommendation that can be made from examining this data is that a power-policy should be implemented on student machines in the University as soon as possible. By, for example, putting machines into standby when student labs close at night†, the annual AEC of student PCs could potentially be reduced to 171.49MW-h, a saving of almost half!
As discussed in the literature review, there are technologies emerging that promise a means of being able to remotely control the power status of devices; CISCO EnergyWise being given as an example. By harnessing a technology such as this, it may be possible to automate the process of powering machines down at the end of each night, effortlessly saving the University large sums of money on its energy bills.
† ACPI S3 mode for 12 hours per night
Monitors are essential companion devices to PC installations throughout the University. Without them, the PCs would lack their main means of communication with users. The monitor and PC are often treated as one combined unit from a retail perspective, but due to their individual power requirements, this report treats them as distinct entities.
In the past, monitors were most always of the Cathode Ray Tube (CRT) type [1]. Now, with more affordable LCD displays available (and the increased picture quality and space saving that these units provide) CRT units are being gradually replaced. The upgrade of monitors has a sound economic backing as well, with a study showing that between monitors of the same size, CRT devices can use up to three times as much power when active [43].
It was assumed that in the University, each PC had an accompanying monitor. Only monitors of the 17” LCD specification could be found in the University at the time of writing, so this type of device was taken as standard. In summary, the total stock of monitors on campus was taken as 2,335 17” LCD units.
To gain exact usage data for monitors over the campus, an extensive usage study would be required. In Roth’s report, figures from various other authors are cited for office hour usage times of various types of equipment. However, this data is geared towards 9am to 5pm office usage and is inappropriate for this study.
Usage data for used in this study was taken to mimic that of PCs: it was assumed that each student monitor was in active use for the 12 hours per day that labs were accessible, while being put powered down into suspend mode or being powered off for the remaining 12 hours. Staff monitors were assumed to be on for ten hours per weekday and in off/standby mode for the remainder of time.
Five different power modes were specified for monitors:
|
Power Mode |
Description |
|
Active |
On and in use |
|
Standby |
Screensaver mode |
|
Suspend |
Screen in sleep mode |
|
Off |
Soft off: Power supply still connected |
|
Unplugged |
Power supply removed from mains |
Table 6.5: Description of Monitor Power Modes
Power consumption measurements were taken with a power monitor for each of these modes.
|
Type |
Active |
Standby |
Suspend |
Off |
Unplugged |
|
17” LCD |
20.25W |
21W |
0W † |
0W |
0W |
Table 6.6: Power draw of monitor in various modes†
Due to the similarity of power consumption values between several power modes, it was possible to simplify power consumption into two modes: Since “Suspend”, “Off” and “Unplugged” modes all provided a measurement of 0W, they are combined into a “Suspend/Off” mode for the purposes of the AEC calculation. Likewise, since the “Active” and “Standby” readings are so similar, they have been averaged and treated as a single “Active” mode with a value of 20.625W.
† Readings gained from monitor in suspend mode were shown as 0W. This suggested that the power monitor used was not sensitive enough to measure the tenths of Watts that were likely being consumed. Assumedly if the reading was 0.5 or above, the power reading would have been rounded up to 1W. Since this implies that the power drawn in suspend mode was 0.4W or less, this reading can be considered negligible and has been taken as 0W.
|
|
Stock |
Mode |
Draw (W) |
Usage (h/year) |
AEC (MW-h) |
|
17” LCD (Student) |
684 |
Active |
20.625 |
4,368 |
61.621560 |
|
Suspend/Off |
0 |
4,368 |
|||
|
17” LCD (Staff) |
1651 |
Active |
20.625 |
2610 |
88.87539375 |
|
Suspend/Off |
0 |
6150 |
|||
|
|
|
|
|
Total |
150.49695375 |
Table 6.7: AEC of Monitors
Calculations estimate that LCD monitors across the University campus consume in excess of 150MW-h per year. Monitors attached to staff machines utilise more power, despite having a dramatically lower usage pattern. This can be attributed to the larger stock of staff machines.
Actual power usage is bound to be less than this as the University’s power scheme requires monitors to power into Suspend mode after 30 minutes of inactivity, making the 12 hour active use estimate per monitor rather generous.
The replacing of older CRT monitors throughout the University (performed prior to and throughout the year of this report, so comprehensively that a CRT monitor could not be found for measurements) has proven to be a wise decision, with Roberson’s study suggesting that power consumption may have been reduced by up to three times!
Interestingly, measurements showed that monitors with a screensaver active utilised more power than those in active use. A speculated reason for this could be due to the brightness of the screen in each mode. The screensaver rendered a black screen with a small logo, meaning that most pixels would have had to be rendered black. The active monitor on the other hand rendered mostly bright colours. This suggests that LCD monitors require more power to render dark pixels than light pixels resulting in systems using dark screensavers being slightly less efficient.
To any paper based company or organisation that uses IT in its day-to-day operations, printers are essential pieces of equipment. They allow forms that have been generated and filled in electronically to be converted to hard copy for filing or correspondence. It comes as no surprise then that printers are common across the University, used not just by administrative staff for functions similar to traditional business requirements, but also for teaching staff and students (who often have high volumes of paperwork to print themselves).
There are many different types of printer available in today’s market, but the University primarily uses two types of printer:
Laser Printers: Tending to be large units for high-volume, high-speed and communal use within the University, these devices utilise xerography (similar to copy machines) but differ in that the image is produced by the scanning of a laser across a blank sheet of paper. Smaller model types are also in use, appropriate for individual use in a staff member’s office, for example.
Inkjet Printers: Almost exclusively small, desktop sized models appropriate for small scale personal printing. To produce their image, minuscule jets of toner are propelled onto a blank sheet of paper. In the University, these are commonly installed alongside a staff member’s PC in their office.
The figures for the stock of printing devices were obtained from data that IT Services was able to provide (attached as Appendix F). As with PCs, only printers determined to be “on the network” are counted.
The distribution of printers in the University network is surprisingly asymmetrical. The printers utilised by staff on the network outnumbers that of the students’ by a factor of four at least. This seems strange until the required applications of each group are considered: Staff printers are likely to be small inkjet printers placed on or near the staff member’s desk whilst student printers will always be large laser printers placed in labs or central locations for communal access.
The 36 student printers are taken to be a mix of medium and large laser printers (the HP Laserjet 4250tn and HP Laserjet 9050n devices, respectively), the former typical of classroom laboratories where individual classes take place and the latter widespread in the larger open access labs, capable of dealing with high volumes of requests.
In order to find out the distribution of inkjet-to-laser printers in the University, a survey was conducted amongst a selection of the University’s staff: 24 members of staff were asked how they typically printed documents in the University.
Detailed results of the survey are included as Appendix G. From these results, it can be established that 11 out of 24 (approximately 46%) of staff used communal laser printers to print their documents. The remaining 13 out of 24 (approximately 54%) staff members said that they had a printer in their office with seven of these being inkjet printers and six of these being laser.
This means that seventeen out of twenty-four staff members use laser printers (71%) and seven use inkjet printers (29%). When applied to the stock of staff printers in the University, these percentages have been applied. This means that of the 159 staff printers total, approximately 46 are inkjet printers whilst the remaining 113 are laser printers.
The inkjet printer chosen to represent the 46 devices present on the network was taken as the HP Deskjet 880c model. A mix of large, medium and small laser printers (the HP HP Laserjet 9050n, HP Laserjet 4250tn and Laserjet P2055d devices, respectively) have been considered to represent the deployment of devices across the staff network. This displays the desktop laser printers revealed in the survey, and medium to large communal devices available for both teaching and administrative staff.
Usage times for student printers were based from the operational time of student labs: The continual supply of power to student PCs inferred that printers received the same treatment. From this, the time for the printers being in off mode was taken as 0.
The power saving capabilities of the laser printers examined by this report allowed them to power themselves down to suspend mode after an hour of inactivity. It was presumed that they remained in this mode for the 12 hours they remained inaccessible to students (labs are typically closed from 9pm to 9am).
The large printer examined (a HP Laserjet 9050n) provides a top speed of 50 pages per minute. This means that even to cope with the printer’s reported load of around 27 pages per operating hour †, the actual time spent in Active mode will remain extremely small each day. The printer’s datasheet (attached as Appendix H) claims the minimum time to deliver a first page is 8 seconds. Generously estimating that each of these 27 pages per hour is a distinct job means that the printer is only active for 216 seconds per hour (or 2592 seconds per operating day). This means that the printer is active for 0.72 hours per operating day (262.8 hours per year), spending the remainder of its time in standby mode. It is presumed that the printers are active enough not to go into suspend mode during the day.
Staff laser printers, being largely communal are expected to have similar usage patterns to student ones. Also, the variance in staff schedules means that communal printers would require relatively flexible access, so the same operating hours have been taken as for student labs.
Individual inkjet printers located in staff offices were assumed to have similar usage patterns to staff PCs and monitors, being turned on for a total of ten hours per weekday.
Thirty minutes of this time was taken as a generous estimate for active printing time.
The remaining time was taken as “off”.
Undoubtedly all staff laser printers would not be of the same size as student communal printers. In order to reflect this and portray fairly the range of device types elicited by the survey, average power draws between small, medium and large laser printer models are used in AEC calculations. It is felt this helps to provide a more accurate AEC value for staff printers.
† Following data taken from a typical printer’s usage report:
Time in operation: Since October, 2007 (approximately 823 days) (Time of writing 9th Feb 2010) Lifetime sheets printed: 270,062 (Usage page) 12 operating hours/day 27.345 pages per hour
Roth’s report recognises several different authors’ attempts to identify the various power states of laser printers. These include three modes (Active/Ready, Standby/Low and Off) by Meyer & Schaltegger and four modes (Active/Ready, Standby/Low, Suspend and Off) by Macebur. Roth’s chosen methodology was to follow Kawamoto et al. (2001)’s approach by taking only two power modes, Active/Ready (to represent a printer that is powered on awaiting print orders) and Off, which represents a printer that is powered off and adding an additional 1W-h per image created by the printer. This method could apply only to laser printers, as Roth acknowledges:
“We did not apply the energy/image methodology to inkjet printers because the 1W-h/sheet energy consumption comes from studies of electrostatic reproduction energy consumption (e.g., Nordman, 1998), which is germane to copiers and laser printers but not the inkjet printing process.” [1, p.62]
As this study did not have usage data for the number of images printer for each individual laser printer, Macebur’s four-attribute approach (Active/Ready, Standby/Low, Suspend and Off) was chosen instead. The inkjet printer chosen did not have a suspend option available to it, so only Active, Standby and Off modes are shown for this device.
|
Power Mode |
Description |
|
Active |
Where the printer is actively printing a document. |
|
Standby |
Where the printer is powered on and awaiting print orders, but is not actively printing. |
|
Suspend |
An approximation of S3 mode where the printer is in a power saving state. |
|
Off |
Where the printer is powered off, but is still connected to the power supply (with phantom load in effect). |
Table 6.8: Power states of printers
Power usage readings for each printer have been taken from datasheets from each model type. (Datasheets for the large, medium, small and inkjet printers used are included as Appendices H, I, J and K respectively). Datasheet readings were compared against actual power draw for one of these models and deemed to be accurate.
In order to represent the power draw of a typical laser printer, the draws of different sized devices were taken and averaged. For inkjet printers, they were treated as one device model.
|
Printer Type |
Mode |
Draw (W) |
Staff AVG |
Student AVG |
|
Active |
Large Laser |
1000 |
750W |
840W |
|
Medium Laser |
680 |
|||
|
Small Laser |
570 |
|
||
|
Typical Inkjet |
30 |
|
|
|
|
Standby |
Large Laser |
205 |
77.667W |
112.5W |
|
Medium Laser |
20 |
|||
|
Small Laser |
8 |
|
||
|
Typical Inkjet |
5 |
|
|
|
|
Suspend |
Large Laser |
36 |
19W |
24.5W |
|
Medium Laser |
13 |
|||
|
Small Laser |
8 |
|
||
|
Typical Inkjet |
n/a |
|
|
|
|
Off |
Large Laser |
0.3 |
0.333W |
0.3W |
|
Medium Laser |
0.3 |
|||
|
Small Laser |
0.4 |
|
||
|
Typical Inkjet |
5 |
|
|
Table 6.9: Typical and average power draws for printers
These average readings were then used in calculating the AEC of all printing devices within the University.
|
Printer Type |
Installed Base |
Mode |
Draw (W) |
Usage (h) |
AEC (MW-h) |
|
Laser Printer (Student) |
36 |
Active |
840 |
262.8 |
7.947072 |
|
Standby |
112.5 |
4117.2 |
16.674660 |
||
|
Suspend |
24.5 |
4380 |
3.863160 |
||
|
Off |
0.3 |
0 |
0 |
||
|
|
|
|
|
Total |
28.484892 |
|
Laser Printer (Staff) |
113 |
Active |
750 |
262.8 |
22.272300 |
|
Standby |
77.667 |
4117.2 |
36.1340746812 |
||
|
Suspend |
19 |
4380 |
9.403860 |
||
|
Off |
0.333 |
0 |
0 |
||
|
|
|
|
|
Total |
67.8102346812 |
|
Inkjet Printer (Staff) |
46 |
Active |
30 |
130.5 |
0.140940 |
|
Standby |
5 |
2479.5 |
0.446310 |
||
|
Off |
5 |
6150 |
1.107000 |
||
|
|
|
|
|
Total |
1.69425 |
|
|
|
|
|
Subtotal |
97.9893766812 |
Table 6.10: AEC of Printers
6 . 5 . 4 . 3 C O N C L U S I O N S
In total, printing devices are estimated to use approximately 98MW-h of power annually. Owing to the larger stock of staff printers, these devices make up the majority of this figure at almost 75%.
When split by type, laser printers make up the majority of the inventory of devices present on the network. Since the University has need for large capacity, high-speed printers, inkjet devices would not be appropriate to fill this role. However as a consequence of using these higher-powered devices, the power consumption of this device type has risen dramatically. In this case, the lower stocks of inkjet printers, and the fact that they use around twenty times less power than small laser printers give an explanation for their relatively small segment of power consumption.
6 . 5 . 5 . 1 B A C K G R O U N D
The University’s network requires a wide array of server computers to store student and staff data as well as provide connectivity from the Internet (in the form of a publicly accessible website and external access to student webmail, amongst others).
The University utilises blade servers in order to provide these services along with several traditional rack mounted server devices. Blade servers are streamlined versions of traditional rack mounted servers, with many of their components being either removed or made more efficient in order to provide a modular design. The advantage of this can be considered the saving of valuable rack space in server rooms, making attached servers more energy efficient.
Rather than having their own distinct power supplies like the traditional rack mounted servers, the individual blades are all mounted into a central chassis which manages power distribution amongst all blades. The chassis itself can accommodate anywhere from one to four power supplies (with more supplies permitting the configuration of load bearing between them, as well as providing redundancy should one supply fail).
6 . 5 . 5 . 2 A E C C A L C U L A T I O N F O R S E R V E R S
Stock of Devices
With data provided from IT Services, the server devices deployed on the University network comprised of the following:
|
Model |
Installed base |
Description/Purpose |
|
IBM X3850M2 |
1 |
Exchange Disaster Recovery Server |
|
IBM X3650 |
2 |
Internet Security & Acceleration (ISA) Servers |
|
1 |
Media Server |
|
|
1 |
CISCO Management Server |
|
|
1 |
Nortel Management Server |
|
|
IBM BladeCenter HS21 |
8 |
Individual blades |
|
IBM BladeCentre H Chassis |
1 |
The 8 HS21 blades are mounted on and receive their power from this chassis. |
Table 6.11: Server Distribution across University
Usage times
It was assumed the University’s servers operated 24 hours a day, 365 days a year (8760 hours) in order to provide constant service to network users.
Roth’s report differentiates between four types of server based on their assumed lifespan. Since this report examines the power usage of servers that are currently in use on the campus, these categories were treated as one. For this study, a separate distinction was made: that between the blade server installation, and the rack mounted servers, since the method of calculating their AEC was slightly different.
The rack servers in use by the University had various rated power supplies on their nameplates. These figures represent the maximum power draw of the device, not its every day power draw. A study performed by Hipp in 2001 showed that the actual power draw of a server is on average, a ratio of 51% of its nameplate value [44]. Datasheets showing the rated power supplies are attached as Appendices L and M.
|
Device |
Installed Base |
Nameplate Draw (W) |
Typical Draw (W) |
Usage (h/year) |
AEC (MW-h) |
|
IBM X3650 |
5 |
835W |
425.85W |
8760 |
18.652230 |
|
IBM X3850M2 |
1 |
1440W |
734.4W |
8760 |
6.433344 |
Table 6.12: AEC of Rack Mounted Servers
Calculating the power consumption of the blade server installation simply required calculating the power usage of the blade centre chassis, since any devices mounted to it (including the 8 HS21 blades) would draw their power from this. Correspondence with IT Services in the University relayed that it was not possible to find out how much power was being drawn from each of the chassis’ installed power supplies. It was also uncertain whether Hipp’s study would be applicable to this device, as his study predated the introduction of the blade server paradigm. However, IT services also implied that it was impossible for each of the chassis’ power supplies to be consistently pulling its nameplate value, so Hipp’s model was tentatively applied. A datasheet showing the power information for this chassis is attached as Appendix N.
|
Device |
Installed Base |
Power Supplies |
Nameplate Draw (W) |
Typical Draw (W) |
Usage (h/year) |
AEC (MW-h) |
|
BladeCenter H Chassis |
1 |
4 |
2900 |
1479 |
8760 |
51.824160 |
Table 6.13: AEC of Blade Server Installation
6 . 5 . 5 . 3 C O N C L U S I O N S
The server computers deployed across the University’s network are estimated to use approximately 77MW-h of power per year. Amongst these servers, it can be seen that the blade server installation consumes the most power, at two thirds of the total AEC. Approximately one third was utilised by the six stand alone rack servers currently in use. The five IBM X3650 servers appear to be fairly low powered in comparison to the single IBM X3850M2, this no doubt being due to the latter server’s increased specifications requiring a larger power supply.
It should be noted that although responsible for the largest amount of power consumption in this scenario, the blade centre chassis has the potential to be more power efficient than the rack servers in use: With a capacity of fourteen blade bays, its chassis will become more efficient compared to these traditional servers as the number of blades installed increases. For this reason, blade servers are recommended for large installations that require the majority of its bays to be filled with blades.
Reason for omission
Hubs are devices that connect several network devices together, treating them as though they were connected to the same cable segment. Hubs do not perform any management of packets that come in through its ports; rather they simply broadcast received frames from all ports except from the one the frame was received on. Hence, each of the devices connected to the hub would be a part of the same “collision domain” where the efficiency of a network can be impacted by devices trying to transmit frames simultaneously. They operate solely at the Physical layer of the OSI model, Layer 1.
Very few (if any) hubs are still in operation in the University’s network. This is largely due to the increasing availability and affordability of access layer switching options which provide more a more efficient way to interconnect devices. The calculation of the hub’s impact on this network has therefore been omitted. Consequently, it is likely that the power footprint of the University’s switching devices will be proportionately larger than shown in Roth’s study.
6 . 5 . 6 . 2 S W I T C H I N G & R O U T I N G
Background
Although switches provide similar connectivity and network functionality to network hubs, they are considerably more sophisticated and as a result offer increased performance. Access-layer switches operate at the second layer of the OSI model, the Data-Link layer, allowing active management of frames that pass through them. One benefit of this is that each connection created between hosts on the switch become part of their own collision domain, ensuring that other traffic passing through the switch will not interfere (or cause collisions with) their communications.
The CISCO hierarchal model defines three “layers” of switching and routing that should occur in a typical LAN. The access layer of the model defines switches which connect directly to end-user devices. These switches always operate at Layer 2 of the OSI model and usually contain 12, 24 or 48 ports in order to service an entire room or floor of a building.
The distribution layer of the model traditionally included LAN based routers and more sophisticated network switches that operate at Layers 2 and 3 of the OSI model. Today, these tasks can be integrated into existing core level devices by installing specialised modules into them, simplifying the topology of the network. For this reason, this layer is considered along with the core switching layer for the purposes of this study. The routing of traffic between different sub-networks and Virtual LANs (VLANs) also happens at this layer.
At the core of the hierarchal model are routers and Layer 3 switches. This layer can be considered the backbone of the network and concerns itself solely with speed and reliability; ensuring packets are transmitted from one portion of the network to the other as fast as possible. No packet manipulation (such as the implementation of Access Control Lists) is performed at this level. As all traffic on the network has to pass through this layer, devices are frequently configured with high redundancy in mind.
There also exists a final class of switch known as the WAN Switch. These devices concern themselves with the delivery of data over large geographical distances and are often used by ISPs to distribute services such as DSL. These devices are not covered in the AEC calculations as it is the Paisley campus’s network alone (and not its links to other campuses) being examined in this report.
AEC Calculation for Access Layer Switches
Stock of Devices
The exact number of access level switch devices deployed across the campus is unknown and unrelated to their AEC value. Instead, a power per port value is required. An approximate number of ports required to service the University’s devices can be deduced simply, since each network-enabled device (PCs and printers) would require one port on an access level switch. The number of ports deployed in this case is the sum of these devices, a total of 2530 ports.
Usage times
In order to provide ceaseless connectivity to the network, it was assumed the
University’s access layer switches operated 24 hours a day, 365 days a year (8760 hours).
Power Draw & AEC Totals
In order to calculate the amount of power consumed by all switches deployed across the network, the amount of power a single port uses must be known. Two main models of switch have been deployed across the University: The Netgear FS728TP and the Nortel 4548GT-PWR. In order to achieve an accurate value for access layer switching, the power-per-port value for each switch was calculated and averaged, before applying to the number of ports deployed across the network.
A datasheet displaying the Netgear switch’s power draw is attached as Appendix O. Power draw information from the Nortel switch was read directly from its supply.
|
Switch Type |
Maximum Power Consumption (W) |
Ports per device |
Power per port (W) |
|
Netgear FS728TP |
225 |
24 |
9.375 |
|
Nortel 4548GT-PWR |
580 |
24 |
24.167 |
|
|
|
Average Power Per Port |
16.771 |
Table 6.14: Power-per-port for Access Layer Switches
|
Ports Deployed |
Average Power per port (W) |
Usage (h) |
AEC (MW-h) |
|
2530 |
16.771 |
8760 |
371.6923188 |
Table 6.15: AEC of Access Layer Switches
Although the figure of 371.69MW-h seems rather high, it must taken into account that these switches are all capable of delivering Power-over-Ethernet and offer the potential to provide expansion possibilities to the University by facilitating easy installation and powering of devices such as VoIP phones, security cameras and wireless access points.
Note that figure cited for number of ports deployed may be slightly inaccurate due to there being an unknown number of wireless access points connected to the network.
AEC Calculation for Distribution Layer Switching, Core Switching & Routing
Stock of Devices
Correspondence with the IT Services at the University revealed that the distribution layer of the University’s switching scheme has been condensed into the core switching layer. The central switch at the core level has a series of line card modules installed in order to provide this function. Therefore, all devices present at the distribution layer are contained within those for the core layer.
The core layer of the University’s switching scheme consists of two Layer 3 CISCO Catalyst 6509-E devices. One of these is kept on cold standby, meaning it has no impact on the network’s power footprint.
The task of routing can be described as the connection of one or more networks or subnetworks. This task was traditionally performed by devices called routers, which operate at Layer 3 of the OSI model. Complex rules and configuration arrangements allow the sophisticated management of traffic that pass through them.
In the University, however, separate devices are not required to perform routing duties. The core layer switch discussed above provides routing functionality between the different VLANs and sub-networks of the University’s network. Two extra CISCO ASA 5580-40 devices are employed to enable routing between the internal network, the Internet and the intermediary Demilitarized Zone (DMZ).
In summary, only one active Catalyst 6509-E and a pair of CISCO ASA 5580-40 devices are active at this layer.
Usage times
Again, to provide uninterrupted network connectivity, it was assumed the University’s access layer switches operated 24 hours a day, 365 days a year (8760 hours).
Power Draw & AEC Totals
The active Catalyst 6509-E switch, much like the BladeCenter chassis, utilises multiple power supplies to increase redundancy and enable load balancing (two 6000W nameplate rated supplies). IT Services was able to provide running power draw figures for both of these supplies, simplifying calculations and providing increased accuracy that estimation might have sacrificed. This information forms Appendix P.
With reference to the CISCO ASA 558-40 data sheet (attached as Appendix Q), these devices are estimated to draw around 800W. The AEC of the ASA 5580-40 pair represents the power required to provide additional routing functionality to the network and is shown on Chart 6.10 as “other routing”.
|
Device |
Installed Base |
Power Supplies |
Nameplate Draw (W) |
Typical Draw (W) |
Usage (h) |
AEC (MW-h) |
|
Catalyst 6509-E |
1 (Active) |
2 |
6000 |
2671.2 |
8760 |
23.399712 |
|
6000 |
2671.2 |
8760 |
23.399712 |
|||
|
|
|
|
|
|
Total AEC |
46.799424 |
Table 6.16: AEC of Catalyst 6509-E
|
Device Model |
Installed Base |
Typical Draw (W) |
Usage (h) |
AEC (MW-h) |
|
CISCO ASA 5580-40 |
2 |
800 |
8760 |
14.016000 |
Table 6.17: AEC of CISCO ASA 5580-40 devices
6 . 5 . 6 . 3 C O N C L U S I O N S
The network infrastructure devices deployed in the University are estimated to use around 418MW-h per year. Comparing the results of the implementation of Roth’s methodology upon network infrastructure devices with his original publication, it is clear to see that network topologies have changed considerably. In Roth’s original publication, he noted large amounts of power consumption from both network hubs and dedicated routers.
By comparison, hubs are non-existent in the University’s network. Cheap switching options (and the improved efficiency of such devices) have provided sufficient to essentially antiquate these devices. Routing, too, is no longer the domain of the dedicated device: One sufficiently powerful Layer 3 switch proves powerful enough to provide all core switching and VLAN routing tasks, something that would previously have taken several separate devices. Even the cluster of CISCO ASA 5580-40s that provide additional routing are not dedicated devices: they also provide additional security features in addition to their routing tasks.
Finally, as in Roth’s study, LAN switches are shown as still having the most impact on the power consumption of the University’s network.
6.6.1 Overview
Comparing the performed study with Roth’s original study performed in 2002, it can be seen that although there are several inconsistencies in the proportions of power consumed, there are also several similarities between the data sets:
6.6.2 PCs & Workstations
Whilst PCs and workstations use just under a third of the total AEC of the five device types in Roth’s report, this study sees an increase in power consumption by these devices. A likely explanation for this is the inclusion of higher specification PCs in this report which although use more efficient power supplies than those in Roth’s report, also use more power in general. Roth’s report also uses a considerable sum of gathered usage data to gauge the operational times of the PCs in his report. In contrast, most of the PCs examined in this report were almost constantly powered on.
6.6.3 Monitors
A vast reduction in power use by monitors and display devices has occurred between 2002 and 2010. In Roth’s study, monitors used far and beyond the most power out of the devices he examined, whilst in this study usage has been reduced to around an eighth of the total AEC figure. The most likely reason for this is the comprehensive upgrade of CRT monitors to LCD screens, resulting in a lower power usage per unit, yielding massive savings in energy.
6.6.4 Printers
The proportion of power consumed by printers in both this study and in Roth’s original study is extremely similar, with both segments weighing in at about a twelfth of the total AEC figure. This suggests that printers have remained relatively unchanged in the eight years that have passed between the studies.
Roth’s figure may be slightly over-stated in comparison, as he also includes impact printers and line printers in his study, both of which had no relevance to this project.
6.6.5 Server Computers
Server computers in Roth’s report use around twice as much power compared to the devices examined in this report. Only one blade enclosure (with eight blades) and six stand alone rack servers were examined in this report, whilst Roth’s deals with an entire country’s stock of server devices. This could contribute to this inconsistency.
The increased efficiency of blade servers and their utilisation of only one power supply combined with these results could suggest that today’s server devices are becoming more efficient and economical than past devices.
6.6.6 Network Infrastructure Devices
The network devices section of the chart shows a massive increase between 2002 and 2010. Devices under this category account for almost three times the proportion of power claimed in Roth’s report. Roth’s lack of investigation into the different varieties of switch device could perhaps account for this deficit, as his report assumed all switch devices were of the common access layer variety, ignoring the impact of the significantly higher powered core and distribution layer switches.
This chapter represents an extensive study on the power usage of network devices in the University of the West of Scotland’s Paisley campus, satisfying Objective 1 of the project. Existing methodologies for power consumption measurements were investigated and the most appropriate one chosen. This was then altered in accordance with the needs of the University’s campus, allowing accurate and representative data to be collected and presented to give a comprehensive breakdown of just how much power is utilised by the University’s network.
With due observation of the results, it is clear to see that there are several ways in which the University can reduce its financial outlay on electricity. Indeed, in one way the University already has: the replacement of aging CRT monitors with power-efficient LCD equivalents will reduce the amount of power consumed by display devices by up to three times.
However, by adopting a power management system to automate the shut-down of computers located in student accessible labs, much more power could be saved. It is for this reason that this report recommends the adoption of such a scheme.
In conclusion, it is hoped that the results of this chapter provide a clear overview of the University’s power expenditure, and that they may be of assistance to its administrators when considering the financial worth of “greening” the campus’s IT operations.
Chapter 6 comprehensively covers the power requirements of the Paisley Campus of the University of the West of Scotland within the scope defined. It is hoped that the data obtained is of some use to the University itself in reducing its electricity expenditure.
Chapter 5 attempts to produce a valid theoretical underpinning for the power requirements of 10-BASE-T Ethernet, creating a mathematical model to calculate the results shown. These results did appear feasible, although seeming surprisingly low. Objective 5 goes on to confirm the validity of the model developed.
Chapters 3 and 4 both explore the effects of “load” on network devices from two perspectives: The effect of a network enabled PC’s processing load on power consumption, and the effect of a network infrastructure device’s network load on power consumption. This objective’s outcomes solely consisted of collecting the data through experimentation, with Objective 5 making comparisons between this data. The successful completion of both experiments can be considered tantamount to its success.
A section of Chapter 5 is dedicated to comparing the theoretical power requirements calculated against those observed through experimentation in Chapter 4. The results gained through the mathematical model developed did match up correctly with the results gained through experimentation, suggesting that they are valid. Although perhaps not as long as any of the other objectives, there was not much more to explore in this area beyond what was said.
Objective 5 consisted of taking the results gained in Chapters 3 and 4 and making comparisons between data gained. The investigation of Chapter 3’s results ascertain that ACPI modes closer to “S6” gain maximum power savings whilst large amounts of processing load can almost double power consumption. Chapter 4 surprisingly notes that the power requirements of data transmission are extremely low, with most power in a typical switch being used in device overheads. Overall conclusions for this objective prove surprising, revealing that controlling the power modes of host PCs may prove to be a more worthy endeavour than making network switches more efficient.
Several problems and concerns encountered throughout this project should be noted:
In summary, the complexity of this project was not realised until the Project Brief had already been submitted. Objectives stated in the document should have been more clearly defined, ensuring that those requiring the most work should be considered advanced. Ensuring the resilience of the Project Brief is advice that could not be stressed more strongly to students considering an Honours level project.
The management of this project can be considered very good, overall. All deliverables were completed before their designated deadlines. Both the Interim Report and the final Dissertation were completed in advance of the submission date.
Experimentation was spread throughout both trimesters, with Chapter 3 being completed before the submission of the Interim Report (admittedly with lower resolution results), and Chapter 4’s experimentation being left until the beginning of Trimester 2. Data gathering for the University power study (Chapter 6) was obtained through a combination of meeting with IT Services personnel, online research and individual investigation into device types, being collected over two trimesters.
All management meetings were attended with minutes being drawn up immediately afterwards, in most cases. On a couple of occasions, the agenda for these meetings were perhaps submitted without much notice. However, on every occasion they were submitted prior to the meeting itself.
Worthwhile advice to future students would be to take note of how long pulling together the final dissertation takes. This project was worked on modularly, in most cases with most objectives consisting of their own documents. Although a highly recommended way to work, one should not underestimate how time consuming formatting and proofreading a document can be.
Chapter 5 of this project perhaps offers an introductory overview of the power consumption of Ethernet technologies. A particularly ambitious Computer Networking student in years to come may wish to expand on it, particularly to explore the impact of increased data rates on power consumption. Certainly, such an investigation may even be worthy of an individual practiced in the discipline of Electronic Engineering. It should be noted that information regarding this area could not actually be located whilst performing the project’s research: Endeavouring to expand this objective could no doubt be considered novel work suited for a level higher than this project.
Based on the foundation lain here, an inter-campus study of the University’s power consumption would be a worthwhile endeavour. Particularly as the University is attempting to lower its energy footprint at the moment, it would be interesting to see whether the power requirements of the campus have decreased. Investigating the power requirements of the links between campuses would also be fascinating.
This project has attained all of its objectives, providing a comprehensive insight into several pertinent issues regarding the power consumption of network devices. The results obtained through some sections of this project have also produced a variety of interesting conclusions, resulting in a range of recommendations that could be made to a host of networking professionals.
Overall, it is felt that the project was performed to a satisfactory standard. The work involved in its completion was extremely rewarding with the entire experience not only being an enjoyable experience, but also having a profound effect on myself as a student.
