topic Hi Sorin, in Intel® Integrated Performance Primitives

ippsFirMR speed

Sorin_Goldenberg — Mon, 26 Oct 2015 06:34:54 GMT

Any idea why the function in a multithreaded program is working at about 1/4 speed than when it works in a single thread?

I am benchmarking the net performance of the functions (it doesn't include any waits, locks, anything)?

Thanks

Soroin

Hi Soroin

Jonghak_K_Intel — Tue, 27 Oct 2015 07:49:00 GMT

Hi Soroin

can we see your test code ?

If multiple threads work on resource that should be accessed by one thread at a time, there could be performance loss for the reason of concurrency control.

- Jon

Hi Sorin,

Igor_A_Intel — Tue, 27 Oct 2015 17:26:19 GMT

Hi Sorin,

could you provide a reproducer? Each IPP release is distributed with IPP PS (perf system) - therefore you can try 2 executable (with -B -r -fFIRMR switches) and compare results from corresponding CSV files:

ps_ipps_mrg_compl_mt.exe (mt==multithreaded) or
ps_ipps_mrg_compl_st.exe (st == singlethreaded)

or run ps_ipps_pcs.exe with multi-threaded dlls/shared objects and single-threaded.

I see good scalability for numIters >= 2K and tapsLen >= 32 (FIRMR has internal criterions on input parameters and mt-code works only for rather large filter orders and number of iterations (also depends on architecture)).

regards, Igor

The ipp is run as single

Sorin_Goldenberg — Wed, 28 Oct 2015 05:04:50 GMT

The ipp is run as single threaded IPP. Multithreading is achieved externally.

What is IPP PS (perf system)?

I work on Windows.

Thanks,

Sorin

Sorin, can you check this

Chao_Y_Intel — Thu, 29 Oct 2015 08:53:08 GMT

Sorin, can you check this article:

https://software.intel.com/en-us/articles/measure-intel-ipp-function-performance

thanks,
Chao