A newer version of this document is available. Customers should click here to go to the newest version.
This timer is provided by the MPI implementation. In general this is simply a wrapper around gettimeofday() and then using it instead of gettimeofday() only has disadvantages: with gettimeofday() Intel® Trace Collector knows that processes running on the same node share the same clock and thus does not synchronize between them. The same information cannot be obtained through the MPI API and thus Intel® Trace Collector is conservative and assumes that clock synchronization is needed. This can be overridden with the SYNCED-HOST configuration option. Another disadvantage is increased overhead and potentially implementation errors in MPI.
If the MPI has access to a better timer source (for example a global clock in the underlying communication hardware), then using this timer would be advantageous.
Did you find the information on this page useful?