Server Processors

[marten]

I wonder if there is a need to consider from an application point of view how CPU cores are allocated to chiplets to optimize performance?

That is, can the performance of an application be improved by considering which CPU cores it uses based on the chiplets (CCDs) of the actual CPU? I am thinking of the Zen3 architecture of AMD CPUs.

If an application for example uses a pool of workers threads pulling jobs from a ringbuffer, shall the workers execute on CPU cores of the same chiplet (CCD) in order to maximize performance of the application?

How can the CPU cores per chiplet be displayed so it is known to which chiplet each CPU core belong?

Thanks

[santosh_zanjurne · Accepted solution]

Hello Marten,

This is very valid and important point in performance measurement of the application / system.  Processes and Threads should mapped to the right cores and should be made to bind-to the core allocated.  There are multiple ways you can do this depending on the type of application you are using.  You can use taskset, numactl, OMP env variables and MPI ranking-mapping.

You can then use AMDuProf to measure the performance of an application and look for remote memory access to see if there are unnecessary remote memory accesses and then plan accordingly.

Regards,

Santosh

[marten]

Thanks for your help.

Best regards

[marten]

Santosh, thanks for your help.

Best regards

[LeiWang1999]

Thanks Reply!

Any research or project focused on this performance optimization progress?