Step 1. Define node configuration
Import the hardware configuration from a file that contains the output of the Intel MPI Library
cpuinfo
utility:
Import from file
Choose file
Browse
Note
cpuinfo
does not contain information about SNC. If you have a configuration with SNC, the NUMA-nodes will be shown as sockets. It does not affect pinning.
Configure manually
Architecture
One physical core in L2
Two physical cores in L2
Number of Sockets
Cores per Socket
Hyper-Threading
Sub-NUMA Clustering
Number of HTs
NUMA-Nodes configuration:
NUMA-node of the same size
Note
Experimental Mode.
Note
By default, the displayed core numbers may not correspond to any real system. Use the Import from file function if an exact match is needed.
Step 2. Define environment variables
I_MPI_PIN_DOMAIN
auto
omp
core
socket
numa
node
cache1
cache2
cache3
cache
manual
I_MPI_PIN_ORDER
default
compact
scatter
spread
bunch
I_MPI_PIN_CELL
unit
core
I_MPI_PIN_PROCESSOR_LIST
Number of MPI-ranks
Show MPI-ranks
Colorize MPI ranks
or use Masklist Editing Mode
Step 2. Masklist Editing Mode:
In this mode, you can manually click on the processor cores to pin MPI ranks to them. As a result, a masklist for
I_MPI_PIN_DOMAIN
will be generated.
After selection, click the "
Next domain
" button to proceed to pinning the next domain or the "
Clear
" button to start again.
You can cancel the pinning of a specific MPI-rank within the current domain
by clicking
on the processor core again.
You can also hold down the left mouse button and move the mouse to select several cores at once.
Next domain
Clear
Exit
Command example:
none
L1-cache
L2-cache
L3-cache
Pinned core
Pinned rank