Reply To: speed up the job in hpc (mpp)

[Reno Genest]

Hello,

After you make sure you run on 32 cores, you can try to bind processes to cores and see if you can speedup the calculation further. I found the following in our knowledge database:

"

By default, an MPI process migrates between cores as the OS manages resources and attempts to
get the best load balance on the system.
But because LS-DYNA is a memory intensive application, such migration can significantly degrade performance
since memory access can take longer if the process is moved to a core farther from the memory it is using.
To avoid this performance degradation, it is important to bind each MPI process to a core.  

Each MPI has its own way of binding the processes to cores, and furthermore, 
threaded MPP (HYBRID) employs a different strategy from pure MPP.


I.  Pure MPP
============

To bind processes to cores, include the following MPI execution line directives according to the type of MPI used.

HP-MPI, Platform MPI, and IBM Platform MPI:

	-cpu_bind or -cpu_bind=rank
        -cpu_bind=MAP_CPU:0,1,2,...   <<<< not recommended unless user really needs to bind MPI processes to specific cores

IBM Platform MPI 9.1.4 and later:

	-affcycle=numa

Intel MPI:

	-genv I_MPI_PIN_DOMAIN=core

Open MPI:

	--bind-to numa
"

Let me know how it goes.

 

Reno.