topic Xeon X5680 slowdown using multithreading in Intel® Moderncode for Parallel Architectures

Xeon X5680 slowdown using multithreading

gilgil — Sun, 24 Oct 2010 14:05:34 GMT

Myxeon has 2 cpus each with 6 cores.
My application performs a cpu-intensive calculation on an image.
The application runs n threads - each with its own image (child buffer) of the same size for k iterations.
I noticed the more threads the higher the time it takes per thread.
I start with 0.83 ms per single runing solely thread and end up with 1.3 per thread with 12 threads.
Setting a thread per core using SetAffinityMask made no improvement.
Another problem rise when using high number of threads - the are a lot more andbigger fluctuations in the time per iteration.
The code itself is mostly sse4 code and the images are of 100X100X3 so there should not be any cache problem.

I would appreciate any idea...

Xeon X5680 slowdown using multithreading

Aubrey_W_ — Mon, 25 Oct 2010 14:40:43 GMT

Hello gilgil,

I'll move this thread to the Threading on Intel Parallel Architectures forum. I'm sure someone there will be able to answer.

Best regards,

==
Aubrey W.
Intel Software Network Support

Xeon X5680 slowdown using multithreading

TimP — Mon, 25 Oct 2010 15:20:50 GMT

Increased time per thread with increasing number of threads is normal, at least when multiple threads go through shared cache. The usual objective is to reduce elapsed time by using more cores.
You may be interested in checking whether the threads which share paths to Westmere cache (cores [0,1], [2,3]) may be less efficient than those for which you set affinity to a dedicated path.

Xeon X5680 slowdown using multithreading

Dmitry_Vyukov — Mon, 25 Oct 2010 15:22:59 GMT

Slowdown can be related to exhaustion of memory bandwidth. What is the size of images in bytes?

Xeon X5680 slowdown using multithreading

jimdempseyatthecove — Mon, 25 Oct 2010 16:46:04 GMT

You might want to chart performance per thread using each possible thread count (1, 2, 3, 4, ... 12)

If you program is performing a large number of writes you may notice a plateau, or at least a drastic change in performance per thread (downwards stair step).

Also, this processor has turbo boost.
Meaning the fewer number of busy cores per CPU, the faster they run.
As you use more cores, (potentially when they get hot), then the boost is turned down or off.

Jim Dempsey