topic Re: performance diff between multi-core and multi-proc? in Intel® Integrated Performance Primitives

performance diff between multi-core and multi-proc?

petemoss — Tue, 19 Aug 2008 18:35:49 GMT

My company is considering building a customized multi-processor machine, where each processor is a 2 or 4 core processor. Will IPP work for us in such an environment? We have only used single 2-core processors before.

Are there any performance issues to be aware of? Should we optimize the app so that each processor has its own memory space?

Thanks
Pete

Re: performance diff between multi-core and multi-proc?

Vladimir_Dudnik — Tue, 19 Aug 2008 18:44:58 GMT

Hello Pete,

IPP work on multi-processor, multi-core systems. And of course you will need to specifically optimize your application to take the most fromsystem with 8 or more available cores (although, your legacy code will also run). The particular optimization technique is very specific for application and amount of data it should process.

Regards,
Vladimir

Re: performance diff between multi-core and multi-proc?

petemoss — Tue, 19 Aug 2008 18:57:38 GMT

Thanks Vladimir.

The app in question would be processing thousands of very large images. It sounds like IPP wont give me any problems.

Pete

Re: performance diff between multi-core and multi-proc?

Vladimir_Dudnik — Tue, 19 Aug 2008 19:27:14 GMT

Yes,IPP itself should not cause problem. It was developed to be thread safe (you can call the same functions in different threads on different data and they will not interfere with each other). And it also was designed to utilize threading automatically on multi-core systems (when you link with IPP DLLs or IPP threaded static libraries). We also provide not threaded static libraries, which can be used in OS kernel mode or when application take care on threading above IPP. For many core systems (8 cores and more) it might be benefitial to implement threading on top of IPP, so application will have full control on threads. I would recommend you to take a look on Deferred Mode Image Processing (DMIP) layer we provide in IPP 6.0 beta. It is binary library built on top of IPP which help efficiently work with large images, keeping application memory working set small enough to fit processor L2 cache. The DMIP also implement threading on top of IPP.

Another good sample is image-tiling IPP sample, which demostrate how you can use IPP to process large images by tiles.

Regards,
Vladimir