Power Efficiency – Rack Mount vs. Blade Servers

HP has just published SPECpower_ssj2008 results for a c7000 blade system, the results when compared to existing rack mount server results for the same benchmark make interesting reading. Before I say anything else, kudos to HP for publishing a power benchmark on blades, now if we can only get IBM, Dell, Cisco etc to follow suit.

The blade system used sixteen identically configured blades to achieve the following result:

ssj_ops @ 100%	avg. watts @ 100%	avg. watts @ idle	ssj_ops/watt
7,210,418	2,783	802	1,877

Each of the 16 blades used a pair of Intel Xeon 5520 processors and 8 GBs of memory. The result is interesting because HP just happens to have published results for an identically configured Proliant DL 380 G6 rack mount server. I’ve multiplied the rack mount result by 16 so we can directly compare power efficiency:

ssj_ops @ 100%	avg. watts @ 100%	avg. watts @ idle	ssj_ops/watt
7,037,296	2,720	1060.8	1,813

From these results we can draw some useful conclusions:

• There is a difference of roughly 3.5% for performance/watt, with blades holding the advantage.
• Idle power consumption for the blade solution is approximately 25% lower
• Peak power consumption for the rack mount solution is approximately 2% lower

Assuming both systems are kept 100% busy 24x7x365, and power is $0.095 kW-hour, then the power costs for the rack mount solution would be $52 cheaper on an annual basis. Assuming the systems are 50% idle, the blades would have $44 dollar annual savings on power. So it looks like power consumption (at least based on these results) shouldn’t be a significant factor if you are trying to decide between blades  and rack mount servers.

Posted by: Nik Simpson

Comments

[Chandler Hall]

Interesting. After all the hype about blades and energy savings, I am a bit stunned by this result. I would have expected the savings to be far greater than this comparison.

At least I know I can upgrade to the latest Nehalem dual processor Xeon's, ensuring that I'll have lots of idle time, which should result in 25% savings on all that power I don't need. :-)

Thanks for taking the time to put this together.

[Matt Lavallee]

The result is interesting, but I think far from conclusive: The DL380 G6 has six PCIe slots, 4 LAN ports, and so on, while blades (individually) have none of these -- there *should* be a much higher accumulated benefit than a few percent, particularly when idle.

I think that the newly-posted result on the DL4x170h, which is essentially a conventional 1U server in a multi-tenant configuration (sharing only RPSes), is far more interesting. By my calculations, this unit bests the c7000 by more than 10% in perf/efficiency, cost 50% less, and takes up two fewer rackmount units.

[Nik Simpson]

Thanks Mike,
I agree, I was expecting to see a lot more of a difference. I looked at the 170h result, and I've a suspicion that the actual server blade in the 170h is probably a variant of the motherboard used in the blade. I'd bet the better power efficiency is because the PSU can be sized exactly for four blades with a specific power envelope, rather than having to cater for a range of different blade sizes and configurations.

[MattL]

The 170h is a sister part of HP's new SL product line. I'm pretty sure these are fresh designs, given that they have IO slots and ports and do not share any bus connectivity (as blades do).

(NIK's comment) - They may have just extended to blade board design a little, just replace the section that has the PCIe mezz card connectors with a layout that includes PCIe slots instead. So my bet is that while the boards are not the same, they are probably close cousins.

I'd side with you on the power efficiency boost, since the PSes scale linearly to the nodes. At the same time, you're also being liberated from the blade chassis overhead and complexity.

With cost/perf/efficiency being so much higher on these nodes, I suspect they'll really take off when they get marketed.

-Matt