SPEC server benchmark spices up efficiency race
The SPECpower_ssj2008 energy efficiency benchmark has existed for less than six months, and already the 1000-point mark has been reached. Fujitsu Siemens Computers (FSC) is reporting that its single processor Primergy TX150 S6 server delivers an average of 1018 server side Java operations (ssj_ops) per watt of power consumed. The previous leader in the field, a dual socket server with two of Intel’s brand new 50 watt Xeons and the DDR2 Registered 5100 chipset, managed 910 ssj_ops/W – and knocked the previous leader, the IBM System x3450 – 854 ssj_ops/W –, into second – now third – spot after just five days at the top.
The SPECpower_ssj2008 benchmark has attracted criticism for measuring server performance based on only one Java application. The results won't necessarily reflect the efficiency when it is used as a web, data, print or database server. In addition, server manufacturers are free to use whatever operating system, Java engine or hardware configuration they wish, so companies naturally select test configurations which show off their products to best effect. Record holder FSC ran their tests with only the primary power supply connected, removing the optional redundant power supply unit and they used only a single SATA hard drive, dispensing with the "World class … redundancy” boasted of in the advertising. Nonetheless, the Primergy TX150 S6 was equipped with the full complement of 8GB of memory and the fastest UP Xeon X3360 – 45nm four core, 2.83GHz, 95 watt TDP –, both of which helped ensure that the BEA JRockit Java engine, which is rarely used other than in benchmarking tests and is optimised for Intel processors, achieved maximum performance under Windows Server 2003 R2 X64.
Despite criticisms of SPECpower_ssj2008, the benchmark is making server manufacturers build more efficient systems. And because all manufacturers are setting up their systems similarly for optimal performance in the benchmark tests, the figures are at least reasonably comparable - if you were to install the same number of additional hard drives in each of the benchmarked servers, the relative comparison between the machines would likely change very little. If the difference in efficiency is large, small differences in hardware configuration are likely to play a relatively minor role. Compared to four year old NetBurst single core Xeons, the latest quad core processor machines offer more than ten times the Java performance per watt.