Releases: efocht/hpcg-ve-bin
HPCG-3.1-v3.1.10ef
This is a binary release of HPCG v3.1 for the SX-Aurora TSUBASA Vector Engine.
Best grid size for one VE that I could find: 56 x 216 x 376.
When running on multiple VEs you should increase nx to at least 72 and scale down the other parameters. For 8 VEs and 64 processes 80x144x376 is a decent choice filling the VE memory quite well.
You must use NUMA for high performance! Switch the system over with the recipe described here.
The reference run preceeding a normal optimized run is taking very long! 11 minutes for the gridsize mentioned above. After the first run of a certain geometry the reference tolerance will be stored in a file called .ref_...
. This file will allow to skip most of the reference run in future runs. Do not manipulate these files!
Install the RPM libhugetlbfs-utils
on your machine and make sure it has a decent number of huge pages configured. The run.sh script calls hugectl --heap
in order to speed up transfers of large VHcall buffers.
Place all files of the release into the directory where you want to run the benchmark. libvhcallVH.so
is essential for running the benchmark.
HPCG-3.0-v3.0.9ef
This is a binary release of HPCG for the SX-Aurora TSUBASA Vector Engine.
Best grid size for one VE that I could find: 56 x 216 x 376.
When running on multiple VEs you should increase nx to at least 72 and scale down the other parameters. For 8 VEs and 64 processes 80x144x376 is a decent choice filling the VE memory quite well.
You must use NUMA for high performance! Switch the system over with the recipe described here. The NUMA setting doesn't persist over reboots of the VH!
The reference run preceeding a normal optimized run is taking very long! 11 minutes for the gridsize mentioned above. After the first run of a certain geometry the reference tolerance will be stored in a file called .ref_...
. This file will allow to skip most of the reference run in future runs. Do not manipulate these files!
Install the RPM libhugetlbfs-utils
on your machine and make sure it has a decent number of huge pages configured. The run.sh script calls hugectl --heap
in order to speed up transfers of large VHcall buffers.
Place all files of the release into the directory where you want to run the benchmark. libvhcallVH.so
is essential for running the benchmark.
hpcg-3.0-ve v3.0.5ef
This is a binary release of HPCG for the SX-Aurora TSUBASA Vector Engine.
Best grid size for one VE that I could find: 56 x 216 x 376.
When running on multiple VEs you should increase nx to at least 72 and scale down the other parameters.
You must use NUMA for high performance! Switch the system over with the recipe described here. The NUMA setting doesn't persist over reboots of the VH!
The reference run preceeding a normal optimized run is taking very long! 11 minutes for the gridsize mentioned above. After the first run of a certain geometry the reference tolerance will be stored in a file called .ref_...
. This file will allow to skip most of the reference run in future runs. Do not manipulate these files!
Install the RPM libhugetlbfs-utils
on your machine and make sure it has a decent number of huge pages configured. The run.sh script calls hugectl --heap
in order to speed up transfers of large VHcall buffers.
Place all files of the release into the directory where you want to run the benchmark. libvhcallVH.so
is essential for running the benchmark.