The new rPerf document from IBM released on the 17th of Aug contains some big rperf numbers for the IBM System P POWER7 795, and due to the dramatic leap in processor core count (64 on a 595, to 256 on a 795) they have paid some thought to the way they display the rPerf figures.

AIX on POWER is one of the best operating systems for scaling into high core counts, but with any OS you will have decreasing returns each time you add a core. This may be insignificant with core counts in single figures but as you get into double or triple figures the overhead becomes significant. If IBM had continued with the ratings that were based on a full system image the results would not represent the workload the IBM Power 795 was capable of when being run virtualised into multiple smaller OS images.

Real world examples of LPARs with greater in size than 64 cores are few and far between, so they have produced a set of rPerfs for the 795 based upon a system split into LPAR's

Unfortunately they have not done this with the rPerf figures for the other models in the range, but I suspect that they will be encouraged to do so by customers wanting to consolidate, who have more modest LPAR core count requirements (apparently the average in the field is 4 cores per LPAR!) and especially now the POWER 770 and POWER 780 are available with 64 cores.

Another confusing aspect to this publication is that they have chosen to compare a different set-up for each CPU speed, i.e.

• 3.7Ghz model has a set of rPerfs is based upon 24 core LPAR's only
• 4Ghz model is based upon 32 and 64 core LPAR's
• 4.25(4Ghz model in turbo core mode) with 16 and 64 core LPAR's

(I guess they have stuck to whole processor book LPAR's for best results)

If you graph the rPerf figures against the core count in the physical box you will notice that the lines are perfectly linear where the LPAR sizes are constant.

..and if you compare the rPerf per core ratios, it is the same from the smallest box to the largest when the LPAR size is constant;

rPerf/core ratio's

• 3.7 Ghz 24 cores/LPAR = 11.39625 rPerf/core
• 4.0 Ghz 32 cores/LPAR = 11.63343 rperf/core
• 4.25 Ghz 16 cores/LPAR = 14.47312 rPerf/core

Which implies that there are no losses when scaling a the the virtualised system. You can however see the per-core loss in scalling a single LPAR when you compare these figures with the 64 core LPAR figures;

rPerf/core ratio

Personally I think IBM should change the format of this document to show more rPerf figures for differing LPAR sizes, rather than different physical box sizes. It is pointless having large tables that just multiply out the rPerf/core ratio according to cores installed where the relationship is linear, it would be much more useful to be able to estimate what rPerf you can expect from an LPAR of x number of cores as this is not a linear relationship.

The rPerf document is a great reference point and it is updated as soon as new kit is announced. It can be found at: http://www-03.ibm.com/systems/power/hardware/reports/system_perf.html

It's a also an interesting record of the POWER chip progress over the past decade!

Read more about the IBM POWER7 range:

View the full IBM POWER range: Click here

POWER6 vs POWER7: taking a closer look at the prices: Click here

View our IBM POWER7 Blogs from our experts: Click here

Speak to one of our IBM POWER consultants by calling us on 0115 951 9666, or leave your details and we'll contact you right away.