Getting IT vendors to agree to any standard, even one that they have a big hand in shaping, is almost impossible. And so it is with the new Energy Star specification for servers, supposedly embraced by server makers to show the energy efficiency of their metal.
The Energy Star specification for servers, which debuted in May, is designed to gauge server power consumption so IT managers can shop for the most energy-efficient machines using an agreed yardstick. In this case, a yardstick established by the US Environmental Protection Agency's popular Energy Star program.
While the Energy Star for servers 1.0 spec has plenty of holes, which we discussed back in May when it rolled out, it is a big step in the right direction. When it is finally completed, perhaps by the October 15 deadline, is hoped that server vendors will adopt a single power-related performance benchmark.
Version 1.0 of the Energy Star spec looks at idle power draw for the iron, which is an important metric to be sure, but when you are measuring the wall, you need a ceiling as well as a floor. Version 1.0 doesn't cover consumption when the server is performing, running loads and tasks.
Yet, Hewlett-Packard was first out the door getting a few of its machines certified under the Energy Star for servers spec, and Dell was pretty close behind.
Neither vendor has opted to use the default power-related performance benchmark, the SPECpower_ssj2008 Java server test that the EPA and various server makers have created in conjunction with the Standard Performance Evaluation Corporation. While the 1.0 spec does not require a specific power-performance benchmark, the default in the docs is SPECpowerssj_2008. The spec and has been doing benchmarks for decades.
HP has put its ProLiant DL360 G6, a 1U machine with two Xeon 5500 processors, and its ProLiant DL380 G6, a more expandable 2U two-socket box, through the Energy Star paces. You can see its power and performance data sheets here.
HP put only one configuration of the 1U box through the test, but did four different configurations of the 2U box. All of the machines were running Windows Server 2008 Enterprise Edition with Service Pack 1, Build 6001.
The 1U box was equipped with two Xeon X5550 processors - four cores each, running at 2.66GHz - and 12GB of main memory, which is six 2GB DDR3 DIMMs. The idle power of the server without a disk drive - since they don't ship standard - was 130.3 watts, but the Energy Star spec allowed this configuration to idle as high as 186 watts.
The ProLiant DL360 G6 was equipped with a "high efficiency" 460 watt power supply, and when running the System Analyser, Diagnostic and Reporting Assistant (Sandra) Engineer 2009 benchmark created by SiSoftware - presumably with at least one disk drive, the report didn't say - the server peaked out at 270.8 watts.
HP tested three different configurations of the DL380 G6 server, and the first was as identical as you could make it to the 1U box tested above. This machine, with the same processors and memory, has a 750-watt high-efficiency power supply. It idled at 139 watts - it had an allowance of 190 watts on the Energy Star spec, given its components - and when running the same Sandra Engineer 2009 test, it peaked out at 260.6 watts.
In this case, the larger box was more energy efficient. Go figure.
Bumping this machine up to 2.8GHz Xeon X5560 processors gave the same 139 watts of idle power, but peak power draw running the Sandra test pushed that up to 270.8 watts - the same as the 1U box.
Just for fun, HP took the DL380 box and put in the 460-watt power supply and only one 60-watt L5520 processor running at 2.26GHz and only 4GB of memory. This machine was allowed to idle at no more than 69 watts based on the Energy Star spec, and squeaked by with 68.7 watts of measured idled power. At peak load running the Sandra test, this server burned 106.2 watts.
HP plans to run the DL160, DL180, DL320, DL350, DL370, and DL385 G5P rack servers - the latter supporting Opteron processors, the rest doing Xeons - and the ML150 tower server through the Energy Star certification process.
Over at Dell, two Nehalem EP-based machines, the PowerEdge R610 and R710 11th generation boxes, have been run through the Energy Star gauntlet. You can see the Energy Star power and performance data sheets for these two boxes here.
Dell only ran two configurations through the test, using Windows Server 2003 as the operating system on the Xeon boxes and picked the Sandra Dhrystone CPU benchmark to stress-test the machines, also from SiSoftware. Daniel Bonds, director of enterprise power and cooling solutions at Dell, didn't want to point fingers as to why Dell didn't use the SPECpower_ssj2008 test, but when pushed eventually said Dell was not happy that the SPECpower_ssj2008 didn't have a standard software stack that would cut down on excessive tweaking and tuning.
The PowerEdge R610 is Dell's most recent 1U rack box, and in the typical configuration with two Xeon E5540s running at 2.53GHz with 12GB of memory and four 2.5-inch SATA drives had an allowance of 218 watts according to the Energy Star for servers 1.0 spec, but this configuration only burned 123 watts at idle and 269.3 watts at peak load running the Dhrystone floating point test. The R610 can be equipped with either a 502-watt or 717-watt power supply, and presumably the smaller one was used in this test of the typical configuration.
The PowerEdge R710 is a 2U box with more expansion room but the same two Xeon processor sockets and with slightly fatter power supplies, at 570 or 870 watts. In a typical configuration with two Xeon E5540s, 12GB of memory, and four disks, the Energy Star spec says it can have an idle power of 224 watts and still be called energy efficient, but this Dell box idled at 122.3 watts. With the Dhrystone test running, it drew a maximum of 266.8 watts.
The current Energy Star for servers spec does not including power ceilings for machines when running a benchmark workload, but is expected to when the specification is updated this fall. Hopefully by then, the SPECpower_ssj2008 software stack can be locked down sufficiently so at least some variant of it can be agreed upon as the default benchmark for the Energy Star ratings.
Once that happens, it will be possible to compare energy efficiency across different makes and models of servers and provide more useful information for data center managers. ®