HPC Resource Kit: Tracing the Execution of MPI Applications with Windows HPC Server 2008
< Go Back
The version of the Microsoft Message Passing Interface (MS-MPI) included in Windows HPC Server 2008 has been paired with the Event Tracing for Windows (ETW) infrastructure in the Windows client and server operating systems, giving application developers the ability to create traces in production environments, create MPI event logs from all processes on all nodes running MPI applications, and tap into "live" event streams.
This article describes the use of ETW in combination with Windows HPC Server 2008. It begins with the four basic steps: adding a "-trace" argument to the command, which launches your MS-MPI application; creating the CPU clock synchronization data for each process; formatting the binary .etl file for viewing and analysis; and copying the formatted event files to a single location and, optionally, merging them into a single, time-correlated log of MPI events on all processes and all nodes.
A sample script to automate the MPI trace process is provided. The paper then explains how to interpret the trace output by reading the text event logs and by using the trace files view in Jumpshot. Finally, the paper covers some advanced tracing topics, such as troubleshooting, tracing applications that are integrated with the Windows HPC Server 2008 Job Scheduler, and tapping into the "live" ETW traces.
