shared High-Performance Computing (sHPC)¶
Last changed: 2023-09-22
This document describes the shared High-performance computing (sHPC) service offering in NREC.
What’s different¶
The sHPC service offering in NREC differs slightly from the normal services. The purpose is to offer flavors for compute or memory hungry workloads and isolate them from impacting workloads in the normal services.
- sHPC workloads are still sharing resources in the same way as in the normal services, albeit with much less overcommit ratios. This ensures that the workloads in the sHPC service are guaranteed a high degree of available resources even though there are no dedicated resources. This leads to better resource utilization for the NREC infrastructure.
- The sHPC service has three flavor sets weighted for compute and memory intensive workloads respectively. Read more about the available flavors belov.
- Continuous uptime is not provided in sHPC service because the instances have their hard drive located locally on the hypervisors. Instances will be taken down in planned maintainance windows, scheduled for the second Tuesday of the month.
These are the key differences in the NREC infrastructure for sHPC instances compared to normal instances:
| sHPC | Normal |
|---|---|
| AMD EPYC processors, 2nd gen or newer. | Various model and generation Intel processors. |
| Slight overcommitment of CPU and memory for best resorce utilization | Resources such as CPU and memory are overcommitted, as workloads usually don’t use more than a fraction of the given resources. |
| Maintainance windows forcing instances to be shut down while the NREC team performs upgrades on the hypervisors. | No scheduled downtime for instances. |
Getting Access¶
Please use the normal form to apply for a project and apply for a shared project, then check the box “Need access to shared HPC” in order to gain access to the sHPC flavors. If you already have a shared project, please use the normal support channels as described on our support page in order to apply for access.
Flavors¶
The sHPC flavors are divided into three categories:
| Flavor Type | Available in Region | Main Purpose |
|---|---|---|
shpc.m1a |
BGO, OSL | Balanced workloads |
shpc.c1a |
BGO, OSL | Compute intensive workloads |
shpc.r1a |
BGO | Memory intensive worklads |
In addition, we have the following disk optimized flavor sets. These are only available in the OSL region, and uses very fast local NVMe storage on the hypervisor:
| Flavor Type | Disk Size | Main Purpose |
|---|---|---|
shpc.m1ad1 |
200 GB | Balanced workloads |
shpc.m1ad2 |
500 GB | Balanced workloads |
shpc.m1ad3 |
1000 GB | Balanced workloads |
shpc.m1ad4 |
2000 GB | Balanced workloads |
shpc.c1ad1 |
200 GB | Compute intensive workloads |
shpc.c1ad2 |
500 GB | Compute intensive workloads |
shpc.c1ad3 |
1000 GB | Compute intensive workloads |
shpc.c1ad4 |
2000 GB | Compute intensive workloads |
The disk optimized flavors are identical to their «m1a» and «c1a» counterparts, except for the OS disk size. The disk optimized flavors are only available by request.
Detailed information about the sHPC flavors:
| Flavor Name | Virtual CPUs | Memory | Access |
|---|---|---|---|
shpc.m1a.2xlarge |
8 | 32 GiB | by default |
shpc.m1a.4xlarge |
16 | 64 GiB | by default |
shpc.m1a.8xlarge |
32 | 128 GiB | by default |
shpc.c1a.2xlarge |
8 | 16 GiB | by default |
shpc.c1a.4xlarge |
16 | 32 GiB | by default |
shpc.c1a.8xlarge |
32 | 64 GiB | by default |
shpc.c1a.16xlarge |
64 | 128 GiB | by default |
shpc.r1a.xlarge |
4 | 32 GiB | by request only |
shpc.r1a.2xlarge |
8 | 64 GiB | by request only |
shpc.r1a.4xlarge |
16 | 128 GiB | by request only |
shpc.r1a.8xlarge |
32 | 256 GiB | by request only |
shpc.r1a.12xlarge |
48 | 384 GiB | by request only |
The therm vCPU refers to physical threads, which is two pr physical CPU core. So, 64 vCPUs translates into 32 physical, multithreaded real CPU cores.