Odd Prime95 benchmark behaviour

[studeimus]

Hi all. I'm trying to understand the performance characteristics of a 20cpu blade node in esxi compared with a regular is desktop machine. using the Prime95 "torture test" all makes sense... using the small fft test, both esxi perf charts and win8 guest VM task manager (resource monitor) show that CPU core utilisation (num cores used) is proportional to the number of torture test threads to run. That is fine.

The part I don't understand is why, when using the Prime95 Benchmark tool (V28.10, build 1) the opposite seems to be true. ie the test starts with 1 reported thread, but esxi perf charts and guest VM resource monitor show ALL cores max'd out. As the test continues it progressively adds threads (up to the final 20 for this node) and each added thread results in ONE extra cpu core becoming idle, such that when it gets to 20 threads, only ONE (or so) cores are being used.

Its as if the benchmark tool is outputting the thread counts in reverse order! Any thoughts?

Stu.

[Prime95]

Originally Posted by studeimus

Hi all. I'm trying to understand the performance characteristics of a 20cpu blade node in esxi compared with a regular is desktop machine. using the Prime95 "torture test" all makes sense... using the small fft test, both esxi perf charts and win8 guest VM task manager (resource monitor) show that CPU core utilisation (num cores used) is proportional to the number of torture test threads to run. That is fine.

The part I don't understand is why, when using the Prime95 Benchmark tool (V28.10, build 1) the opposite seems to be true. ie the test starts with 1 reported thread, but esxi perf charts and guest VM resource monitor show ALL cores max'd out. As the test continues it progressively adds threads (up to the final 20 for this node) and each added thread results in ONE extra cpu core becoming idle, such that when it gets to 20 threads, only ONE (or so) cores are being used.

Its as if the benchmark tool is outputting the thread counts in reverse order! Any thoughts?

Stu.

Try version 29.2. Benchmarking has been completely overhauled.

In short 28.10 made a feeble attempt to mimic "real world" conditions by loading up the other cores with compute bound jobs. Years ago you would run a prime95 benchmark and see the one core case be very fast due to turbo boost. Then you'd run prime95 and see nothing like that performance when each core was running the same FFT length.

Anyway, prime95 is now-a-days limited by memory bandwidth, so the attempt to neutralize turbo boost no longer is relevant. Thus, version 29.2 no longer spawns these CPU-bound tasks.