Using OpenMP I divided a simple algorithm in different threads but the execution time is increasing drastically. This may be due to all the threads running on same CPU core. I am aware that if my CPU is dual core or quad core then assigning number of threads more than number of CPU cores will not help much. but even with two threads the execution time is increasing.
Hi Avipriyo,
you can use GOMP_CPU_AFFINITY variable and set which thread to use which CPU core. For example, if you have 16 cores, and use only 4 threads, you can set to execute to cores 0-3, or 0, 4, 8, 12.
Try to analyze the cache line utilization in parallel and sequential execution, where a huge trap could happen. If you divide an array of 100 elements 0-49 and 50-99 to corresponding two cores will be much faster rather than to execute dynamically (odd to Core0, and even to odd 1).
Latter case will generate more cache misses, especially if the cache line is smaller and the elements are 4 or 8 bytes.
Hi Avipriyo,
you can use GOMP_CPU_AFFINITY variable and set which thread to use which CPU core. For example, if you have 16 cores, and use only 4 threads, you can set to execute to cores 0-3, or 0, 4, 8, 12.
Try to analyze the cache line utilization in parallel and sequential execution, where a huge trap could happen. If you divide an array of 100 elements 0-49 and 50-99 to corresponding two cores will be much faster rather than to execute dynamically (odd to Core0, and even to odd 1).
Latter case will generate more cache misses, especially if the cache line is smaller and the elements are 4 or 8 bytes.
Hi Sasko Ristov sir,
Thanks a lot for your answer. Yes GOMP_CPU_AFFINITY as well as KMP_AFFINITY variable is also another way I found in Intel developer section. And I will keep in mind about the cache line utilization as you suggested. Thank you once again.
Hi Avipriyo,
you can check my previous research in this direction:
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?reload=true&arnumber=6596257
Hi,
Also remember that there is an overhead for launching and joining the threads, so if the loop you are trying to parallelize is not big enough, you will get slowdown instead.
Hi Farshid sir,
Thank you for answering. Yes I am aware of the overhead of launching and joining the threads. I am bit new to this and i am learning so I am implementing this in small algorithm as for now and later I plan to implement it in heavier algorithms and examine the difference in efficiency.
Hi,
Try increasing the problem size to see the effect of parallelization. This may help.
If you don't set any CPU affinity explicitly the threads should use all cores. From my experience it is faster (at least on a normal computer that is running other Software and a desktop environment at the same time) to let the operating system to do the scheduling to the CPU cores.
Caching can easily be a problem. Try setting different types of schedules and use a chunk size a lot larger than 1 as well. Furthermore, it is important that you have no critical sections in you parallel part. These will serialize the execution of your code with a large overhead.
hey Avipriyo
You want to assign specifically some threads to some cores so that execution time decrease? Maybe this won't work!
have you considered the decomposition of data for your problem correctly? I mean do you break your data in a way that your cpu cores doesn't have to wait for each other?
Did you considered communication overhead? this communication time will always fight back with the time decrement!
Hello Simon sir,
Thank you for answering. Yes you are right. According to my problem, letting the OS do the scheduling is giving much more efficiency than explicitly dividing. I am thinking of applying it on much larger problem where the overhead of processing is much more than the thread creation and communication. Then explicit threading might be beneficial.
- Badges
- Science method