rPerf


 
Thread Tools Search this Thread
Top Forums UNIX for Dummies Questions & Answers rPerf
# 1  
Old 10-25-2012
rPerf

PHP Code:
Machine=IBM,9119-595 MHz=2102 Rounded-MHz=2100 CPUs=8                   
CPUType
=PowerPC_POWER5                                                  
lookup IBM
,9119-595_2100_8                                              
matchup 16 108.13 24 155.51 32 202.88 40 243.46 48 284.04 56 324.61 64  
365.19                                                                  
calculate cpus
=8 from 16 108.13                                         
54.06 rPerf estimated                                         
                                                                        
Machine
=IBM,9117-MMB MHz=3108 Rounded-MHz=3100 CPUs=12                  
CPUType
=PowerPC_POWER7                                                  
lookup IBM
,9117-MMB_3100_12                                             
matchup 16 165.30 32 306.74 48 443.06 64 579.39                         
calculate cpus
=12 from 16 165.30                                        
123.97 rPerf estimated 
Ok, it says pSeries 640 as a baseline of 1.0, which doesn't explain much.
I have 54.06 rPerf for 9119 and 123.97 rPerf for 9117 for these 2 servers.

I don't think 123.97 rPerf for 9117 is 2.3 ( = 123.97 / 54.06) times faster or more capable than 54.06 rPerf for 9119.

What will be the best to estimate 9117 is how many % faster or more capable
than 9119?

Please advise. Smilie
# 2  
Old 10-25-2012
Let's just look at cpu "power", not hyperhtreading or other factors.

Machine=IBM,9119-595 MHz=2102 Rounded-MHz=2100 CPUs=8
2100 * 8 = 16800

Machine=IBM,9117-MMB MHz=3108 Rounded-MHz=3100 CPUs=12
3100 * 12 = 37200

37200 / 16800 = ~2.212

So a priori, the 2.3 looks valid to me. I DO NOT know IBM architecture, but in Sun and HP boxes there are differences in the way cpus can be threaded, caching, bus throughput, disk vs SAN and so on which could give +/- values to performance.

Therefore, why do you suppose those numbers are not representative?
# 3  
Old 10-25-2012
You might find that for one CPU on an idle 9119 might outperform a CPU on the 9117 if the other CPUs were active, because of contention for shared resources like caches, buses, DRAM, DRAM rows being RAS closed. Sometimes jobs run at disk or lowest ram speed because of large working sets and random access. Benchmarking is an art, half the time the opposite of writing a good program! One programmer trying to max out the parallel floating multiply on one SPMD system powered it down for high current. He groomed it so every cycle he could was useful, data just filled every space and all was used, which had never happened before, never in real life. Smilie
 
Login or Register to Ask a Question

Previous Thread | Next Thread
Login or Register to Ask a Question