CLUSTER INTERCONNECT BENCHMARK RECORD
Yenra | September 7, 2026
Guide: https://yenra.com/high-speed-cluster/

HOW TO USE
Reserve an appropriate test period. Keep the application input, build and process placement consistent. Verify output, retain repeated measurements and label background load. Separate theoretical link rates from measured useful throughput and record units and direction.

RECORD
Copy the repeatable task/run fields for each observation. Keep the units shown.

1. Date/operator; application/build/input; expected output and check
   ____________________________________________________________
   ____________________________________________________________

2. Nodes/CPU/memory; adapters; switch topology; firmware; link rates (Gb/s)
   ____________________________________________________________
   ____________________________________________________________

3. MPI and benchmark/version; ranks/node; process binding; host or device buffers
   ____________________________________________________________
   ____________________________________________________________

4. Test type; message bytes; pair placement; shared traffic
   ____________________________________________________________
   ____________________________________________________________

5. Individual latency measurements (microseconds); summary and variability
   ____________________________________________________________
   ____________________________________________________________

6. Individual bandwidth measurements (GB/s or GiB/s, state direction)
   ____________________________________________________________
   ____________________________________________________________

7. Application run times (seconds); output checks; baseline and changed configuration
   ____________________________________________________________
   ____________________________________________________________

8. Bottleneck interpretation; complete upgrade inputs; decision; retest trigger
   ____________________________________________________________
   ____________________________________________________________

WORKED EXAMPLE (ILLUSTRATIVE, NOT MEASURED RESULTS)
Fictional decimal-byte model: 5 microseconds + 1,000 bytes / 2 GB/s = 5.5 microseconds. For 100 MB: 50.005 milliseconds. Nominal 25 Gb/s / 8 = 3.125 GB/s before overhead. Eight 25 Gb/s ports / two 50 Gb/s uplinks = 2:1 aggregate ratio for crossing traffic.

SOURCE CONTEXT
Use the linked guide for explanation and these exact-source documents for
model/version-specific behavior. Recheck current documentation before a change.
https://mvapich.cse.ohio-state.edu/benchmarks/
https://www.mpi-forum.org/docs/
https://hpc.llnl.gov/documentation/tutorials/introduction-parallel-computing-tutorial