MPI Cluster¶
The MPI cluster is optimized to support parallel workloads running across many nodes at once. High-speed networking enables low-latency communication between processes, making the cluster ideal for tightly coupled codes using the Message Passing Interfaces or other distributed computing frameworks.
Jobs use a minimum of two nodes
The MPI cluster allocates at least 2 nodes per job. If you regularly run single-node workloads, submit to the SMP cluster instead. Running a single-node job here still reserves the unused node(s), and those resources are counted against your Resource Allocation usage.
Specifications¶
Nodes are grouped by partition, newest hardware first.
| Partition | Nodes | CPU | Cores/Node | Mem/Node | Mem/Core | Scratch | Network | Node Names |
|---|---|---|---|---|---|---|---|---|
| ndr | 18 | AMD EPYC 9575F | 128 | 1.5 TB | 12 GB | 3.2 TB NVMe | NDR-200 IB | mpi-n[136-153] |
| mpi | 136 | Intel Xeon Gold 6342 | 48 | 512 GB | 10.7 GB | 1.6 TB NVMe | HDR200 IB | mpi-n[0-135] |
Related¶
-
What a job costs
How Service Units are calculated — remember an MPI job bills for every node it reserves.
-
Request resources
Set nodes, tasks, cores, and time for a multi-node batch job.
-
Limits & priority
Per-group CPU and memory limits, QoS walltimes, and how priority is computed.