- Integrated Systems
- About Us
- Integrated Systems
- About Us
08-01-2012 01:54 PM
A good discussion on 130W CPU's and Power Advisor circuit sizing vs OA Power Allocated.
Richard was looking to test higher wattage CPUs with a customer:
I’ve a generic question relating to the expected behaviour of a c7000 chassis with n+n redundancy when it is populated with something like 16 x BL460c Gen8 with the high performance 130W E5-2680 CPU’s and the circuit sizing (Power Advisor) exceeds power supply capability (n+n).
One of my customers is looking to pre-populate ‘pools’ of hardware resources based on these CPU’s with either 16 x 32GB, 16 x 16GB or 16 x8GB memory population (so quite high specification blades and therefore high power consumption). Based on these hardware specifications, the Power Advisor utility only allows me to fully populate a chassis with 16 blades using the 16 x 8GB configuration whilst retaining a n+n power redundancy. With the 32GB DIMM’s I can put 13 servers and with the 16GB DIMM’s then it’s 15 servers. It’s the circuit sizing figure that is above the n+n 2400W PSU rating, the 100% utilisation figure never peaks above around 6100W.
Having never had the issue of not being able to fully populate a chassis before I’m most interested in how the chassis would be expected to behave compared to the circuit sizing figure from the Power Advisor.
Is the c7000/OA Power Allocated figure directly related to the power figure I’m seeing in Power Advisor and therefore I assume if the customer deploys n+n with the 32GB x 16 server config they would insert the 14th blade and it wouldn’t power up, my assumption is this is the expected result? If my understanding is correct this is to limit exposure to circuit power trip issues should the OA’s fail meaning any power capping or static limits set would be overridden?
We’re trying to navigate the ‘cleanest’ deployment solution for the customer and the options so far look like:
1) Implement a process that limits the number of blades per chassis depending on configuration – customer needs a repeatable and standard approach and this looks untidy with ‘spare’ bays and they don’t think it ‘feels’ right, they may get used to it over time though, think it was just a shock to them that they would have to consider it!
2) Change power redundancy to n+1 – I’m less comfortable with this, a couple of issues beyond the obvious higher risk for PSU failures, the customer deploys 3 phases enclosures today and that would need to change and they would lose power feed redundancy (not sure how the chassis would react if a feed did go, would 3 PSU’s go offline therefore bring the whole chassis down?)
3) Try to find a mix of blade configurations in a chassis that is under the Circuit Sizing – proving a challenge beyond the 16 x 8GB DIMM config
Any real-world input would be welcome if anyone has had customers testing these high power CPU deployments.
Input from Monty:
Power Advisor is designed based on lab power consumption measurements to calculate how many power supplies are required based on a configuration for each blade and the rest of the enclosure.
The Onboard Administrator is designed to take advantage of actual max power profiling of each ProLiant server blade – to ensure that servers that are granted power will stay powered if the enclosure loses power redundancy.
OA Power Allocated to a ProLiant server blade is based on two pieces of information:
- Server blade FRU report of absolute maximum power for that server blade
- Measured max power information determined by the BIOS during POST
The server blade FRU max power information can be hundreds of watts higher than the actual measurement of power by the BIOS.
Variables in server max power consumption include:
- CPU type and frequency – which varies by individual CPU
- Main logic board power consumption
- DIMM size and number
- Number of mezzanine adapters
- Number and type of installed disk drives
The OA will permit a server to power on if the requested power (from 1 or 2 above) is less than Power Available.
The OA cannot derate the server power request based on workload – as that server blade will draw the maximum power when it is not idle.
Average server power consumption is the sum of idle power consumption times the time the server is idle plus the max server power consumption times the time the server is not idle, divided by the total time period.
Any other comments or questions?