HP ProLiant Power Efficiency

[chuckk281]

Stephen's customer had a power related question:

*********************************************************************************************************

Current max. power requirement for a HP chassis w/ 16 blades (G6 model x 24 GB RAM) is 3947 watts at 100% utilization

I want to know how much peformance trade off if we limit the power requirement at 3000 Watts.

Thanks,

--------

I have a question about calculated performance when power capping is deployed at the c7000 enclosure level.  If I follow the chart that I’ve included I would assume no performance impact until roughly 80% utilization is met. When the utilization exceeds 80% and proceeds to hit 100%, I see a 5% drop in CPU processing power.

So how much performance is lost when we cap a c7000 to 3000 Watts when it was built to 3947 Watts? That represents a 76% power cap.  According to the chart in the first slide, I would “almost” assume there would be no tradeoff.  Could someone comment on this please?  I would like some engineering material to help provide guidance to the customer if at all possible.

**************************************************************************************************************

Monty jumped in:

**********************************************************************************************************************

Not a simple answer to your question – the power capping algorithm results in different performance impact for different workloads – and is best characterized by measurements done with the customer workload.

The iLO 2 with Advanced license adds the ability not only to perform power capping, but also to view server power at 20 second intervals on the power history graph.  Since the graph includes both power min/max/average and the power cap during each measurement interval, you are able to see what happens to server power while the power capping algorithm is active.

You would need some other measurement to correlate the power capping with application throughput.

With the Onboard Administrator Enclosure Dynamic Power Capping feature – the actual power cap for each supported server blade is adjusted to protect the circuit breakers for the entire enclosure.  This means that if some of the servers are at low CPU utilization – they may be capped much lower than servers that are experiencing high CPU utilization.

Severe power capping of an entire enclosure will result in overall lower peak performance if all blades are at high CPU utilization, but may not show significant impact to light workloads.

Bottom line – there is no linear correlation between power capping and server performance.

*****************************************************************************************************************

David also chimmed in:

*************************************************************************************************************

Stephen:  To expand on Monty’s point and based on the scenario you propose below, it’s very possible that your 3000W power cap may actually never result in having to apply low-power p-states to any of the blades (or at least only in rare occasions).  That, of course, will depend on the workloads running on the blades and the coincidence of peaks among the workloads.  That’s something we often forget:  while aggressive power-capping may require the use of low-power modes (I would characterize aggressive as around 50-55%), a relatively lax 76% as you detail below will probably let all of the blades run at a standard p-state until about half of the workloads simultaneously start going past 75% utilization.  (my reasonable estimate; don’t turn around and quote the figures as a feature!)  My point is that Dynamic Power Capping doesn’t have to equal any performance sacrifice, but it all depends.

*******************************************************************************************************************

To which, Stephen replied:

****************************************************************************************************************

Thank you very much Dave and Monty,

Although this is in line with what we’ve been working with for years it’s still very important and helpful to review it now. Not seeing any new data on the impact to performance caused me to reach out.  The ultimate answer is depends; on work load intervals and application behavior.  I think there’s some really strong rational points here to get the customer to move forward.

******************************************************************************************************************

Let us know how you would handle this.