In many HPC applications, a single slow node can reduce the performance of the whole cluster 50% or more! With so many factors that can slow down network performance, it's difficult to ensure your cluster operates at full potential.

Microway is now offering a free MPI Link-Checker performance test of your compute cluster! By running a short MPI job and sending us the results, you can determine how well your cluster is operating.

Microway will keep your results completely confidential. We list our data gathering in the FAQ.

• Download Microway's MPI Link-Checker test application: mpilc.c
  
  
• Compile the C application:
  

  mpicc mpilc.c -o ~/mpilc

  
  
  
• Distribute the mpilc executable to all nodes in the cluster. If your home directory is a network filesystem shared by all nodes, this will not be necessary. On a Microway cluster, you may distribute the file with this command:
  

  rcpf ~/mpilc ~/

  
  
  
• Execute mpilc on all nodes in your cluster (you will need to replace the 'x' below with the number of nodes). To obtain the most useful measurements of network performance, you should run the MPI program with one process per node, NOT one per processor. The application will write all data to stdout, so the output should be redirected to a file.
  

  mpirun -np x ~/mpilc > link-checker.data

  
  
  
• Upload the output file to Microway for processing and analysis. Microway will provide a plot detailing both network latency and bandwidth. If any issues or slowdowns are detected, we will provide a diagnosis. Finally, we'll provide a performance comparison between your cluster and our clusters.
  
  

• Will the MPI Test work on my cluster?
  Our MPI Test was designed to be compatible with all recent versions of MPI and most C compilers. If you encounter compilation or runtime errors, please report them to sales@microway.com with MPI TEST in the subject line.
  
  
• How long does the MPI Test take?
  The test runs for approximately 1-2 seconds for each compute node. This means a 30-node cluster will test for about 60 seconds.
  
  
• What data is gathered by the test and sent to Microway?
  Our MPI Test gathers network performance data relating to response times and total data sent between nodes. The only other information collected by the test is the date and the name and MAC address of each compute node (for identifying faulty nodes).
  
  
We will not share your name, contact or specific cluster information with any other parties. Although we may release general cluster performance information (e.g., "Of all InfiniBand clusters tested, 10% exhibited performance issues."), we will not release identifying information.