Run Cluster Validation Suite (CVS) network and collective communication suites#
2026-10-23
3 min read time
InfiniBand performance, RCCL collectives, and MORI RDMA benchmarks validate cluster interconnect before distributed workloads.
Suite |
How to run |
Config reference |
|---|---|---|
InfiniBand (IB Perf) |
InfiniBand (IB Perf) test configuration file for Cluster Validation Suite (CVS) |
|
RCCL |
ROCm Communication Collectives Library (RCCL) test configuration files |
|
MORI (RDMA) |
MORI network configuration file reference for Cluster Validation Suite (CVS) |
See also Run Cluster Validation Suite (CVS) test suites on AMD Instinct GPU clusters for common cvs run flags and workflow.
Test results#
Sample output snapshots from an RCCL performance run.
Important
The numbers in these sample reports and screenshots (bandwidth, latency, and every other metric) are illustrative sample values, not final tuned results. They are shown only to illustrate the report layout.
HTML test report#
The pytest HTML report lists each RCCL perf test and links to the RCCL Run Deck.
The report shows the environment block (CVS version, cluster file, config file), links to the RCCL Performance Report and Run Deck, and a per-test result table covering host info collection, firewall checks, and each collective parametrized by algorithm and protocol.
Run Deck report#
The static Run Deck HTML summarizes the run and links to the interactive viewer.
The run card shows the cluster topology (nodes, MPI ranks, local ranks), the collectives tested, message size range, NIC model, threshold mode, and links to the pytest report, run log, cluster file, and config file.
Run Deck interactive viewer#
The interactive viewer adds series and search filters, charts with legend swatches and reset zoom, and a paginated full results table.
The filter bar narrows results by collective and searches any column value. It shows the total row and series counts and instructions for scroll-to-zoom and drag-to-pan on the charts.
The bus bandwidth chart plots GB/s against message size for each collective, with one color-coded series per collective and dot markers at each measured point.
The full results table lists every collective and message size with the measured bus bandwidth, algorithm bandwidth, and time in microseconds.