Mellanox ClusterKit- Single Application evaluating your InfiniBand Cluster Health

Mellanox ClusterKit is an MPI-based application for evaluating your InfiniBand cluster health at multiple levels.

The MPI-based application provides multiple benchmarks:

• Network-wide- Random/order ring performance
• Per Node-System Memory bandwidth
• Pair-wise Network tests- Latency, Bandwidth

GPU tests (bandwidth and latency):

• GPU to Host memory bandwidth
• GPU to remote GPU (pair-wise, GPU0 -> GPU0, GPU1->GPU1, GPUn->GPUn)
• GPU to remote GPU (neighbor test, each GPU talks to each GPU)

Per rack- Collectives (Barrier, Bcast, Allreduce)

Main Features:

• Runs multiple benchmarks as a single MPI job
• Automatic results reporting and visualization, ready for analysis

Differentiators:

• Per node/rack reports

Example: Pairwise Latency Matrix

• Latency between node pairs
• Raw Data from a subset of Summit nodes (Top 500 #1 cluster)
• Data Includes switch and cable latency
• Topology is clearly visible

Interested in seeing more?

Contact us today and ask for a live demo