Intel® oneAPI Threading Building Blocks
Ask questions and share information about adding parallelism to your applications when using this threading library.
Announcements
Welcome to the Intel Community. If you get an answer you like, please mark it as an Accepted Solution to help others. Thank you!
For the latest information on Intel’s response to the Log4j/Log4Shell vulnerability, please see Intel-SA-00646

32 thread limit on Windows?

e4lam
Beginner
523 Views

We've had user reports that our software (which is based on TBB) is only able to use up to 32 threads on AWS g4dn.12xlarge and g4dn.16xlarge processor instances (with 48 and 64 cores respectively). From information they gave us they said:

OS: Windows Server 2019 Datacenter x64

Processor: Intel Xeon Platinum 8259CL @ 2.50 GHz

We failed to reproduce the issue on  AMD Ryzen Threadripper 2990WX 32-Core and AMD Ryzen Threadripper 3990X 64-core processors but we do not have access to these expensive Intel CPUs.

From further testing, the user said that they also encountered the same problem on a different piece of software that is also running Intel TBB 2019. So I'm reaching out here to see if anyone else has run into this? Or perhaps the TBB team has access to these processors for testing?

 

 

0 Kudos
5 Replies
e4lam
Beginner
511 Views

PS. It seems to only occur when it's on an Amazon AWS instance that has 2 sockets. 

RahulV_intel
Moderator
486 Views

Hi,


Thanks for reporting your issue. Did you try it out with the latest oneAPI release?


Does the issue occur only with a specific software application or any application in general?



Thanks,

Rahul




Mariya_P_Intel
Moderator
471 Views

Hi Edward,

Please follow Rahul's recommendation and check if the issue exists or not when using TBB 2021 from the latest oneAPI release https://software.intel.com/content/www/us/en/develop/tools/oneapi/base-toolkit/download.html 

And to give us more details on the issue, please run the following code to see the number of threads in the problem environment and tell the result:

 

#include <iostream>
#include "tbb/task_arena.h"

int main() {
    tbb::task_arena a;
    a.execute([&] { std::cout << tbb::this_task_arena::max_concurrency() << std::endl; } );
}

 

Thanks, Mariya

Mariya_P_Intel
Moderator
413 Views

@e4lam Hi Edward,

There was no feedback from you for two weeks.

I assume that recommendation to use TBB 2021 solved the case and stop monitoring this thread.

Thanks, Mariya

 

e4lam
Beginner
406 Views

Sorry, for the slow reply. But we could not figure out the issue as we ended up spinning up some AWS instances of the same type and could not reproduce. As far as the client could tell, it was something in their OS that would not give affinity to all the cores. But that they could see that it was not given in the Task Manager for the process. They could enable it manually though in the Task Manager once the process started.

Reply