Hi Joe,
Apologies for any confusion. There are currently 3 public image families available for use with Slurm-GCP:
- schedmd-slurm-20-11-4-hpc-centos-7
- schedmd-slurm-20-11-4-centos-7
- schedmd-slurm-20-11-4-debian-10
The "schedmd-slurm-20-11-4-hpc-centos-7" image is based on our new
HPC VM Image, and includes our
HPC Best Practices Tunings. Today the HPC VM Image disables hyperthreading by default. If you use the "schedmd-slurm-20-11-4-centos-7" image, that will be based on Standard CentOS 7, and will have Hyperthreading enabled by default. You'll also need to set "image_hyperthreads" in your tfvars file to "true". We're updating and improving our documentation around this as we speak, based on user feedback.
We have a feature in testing in the Compute Engine service that will allow customers to disable Hyperthreading at the hypervisor level by specifying the number of threads per core , rather than the Guest OS level as
are outlined today. Once that feature is available the "image_hyperthreads" variable will change, and will call the Compute Engine API to enable/disable hyperthreading as specified. The HPC VM Image will also stop disabling Hyperthreading at the OS level, and direct users to use the Compute Engine API feature.
We'll make an announcement to this group when we've updated the HPC VM Image, and when we've added full support into Slurm for disabling Hyperthreading at the hypervisor level.