Last month, I blogged about HP’s recent announcement of SPECpower benchmark results for blade servers (see here). If you don’t have time to go back and read the original post, the gist of it was that there was very little difference in power efficiency between HP’s blades and their rack-mount servers which contradicts one of the key marketing messages for blades.
Needless to say, HP’s blade marketing team were concerned and promised to investigate further, which they did. As it turns out, the benchmarking team at HP were very aggressive in optimizing power consumption for the rack-mount system, optimizations included:
The rack-mount system was configured with a single power supply, rather than a redundant configuration which is more popular with customers. This has a major impact on power efficiency, because redundant power supplies share the load, so the maximum utilization on a power supply in a dual configuration is never going to be greater than 50%.
The problem arises because of the way SPECpower benchmark works, generating a series of workloads, some of which barely register on a modern server. With these loads, power consumption on a single power supply can easily drop into the 25-30% range (i.e. if it’s a 500 watt power supply it could draw as little as 125-150 watts). With a dual power supply system, the load on each supply would drop into 12.5-15% range which is very bad for power supply efficiency. Modern power supplies under loads from 20-100% are very efficient, typically in the 90% or better range, but efficiency drops very steeply once you fall below 20% utilization. It’s this loss of efficiency that HP was trying to avoid on the rack-mount server by only equipping it with a single power supply.
The way the SPECpower benchmark is written allows for unnecessary features like network interfaces (yes, I know you have lots of servers with no network interfaces) to be turned off in the benchmark.
Different Benchmark Settings
Despite the fact that the blade and the rack-mount server were identical from a hardware perspective, the settings used for the benchmark varied widely between the systems. For example, different JVM command-line optimization and heap allocations were used for the rack-mount and blade server runs of the benchmark.
So while the numbers appear to be “apples-to-apples” in reality the rack-mount result represents artificially high power efficiency that you probably will not see in the real world.
The net result is that the SPECpower benchmark is of little use when comparing power efficiency between different server types and configurations because the vendors can run non-standard configurations, make optimizations that would never occur in the real world, and are not consistent about the way they run the benchmark from one result to the next, even between two benchmarks results from the same company!
Posted by: Nik Simpson