Intel® Trace Analyzer and Collector User and Reference Guide

ID 767272
Date 3/31/2023
Public
Document Table of Contents

Parameter Checking

(LOCAL:MPI:CALL_FAILED)

Most parameters are checked by the MPI implementation itself. Intel® Trace Collector ensures that the MPI does not abort when it finds an error, but rather reports back the error through a function's result code. Then Intel® Trace Collector looks at the error class and depending on the function where the error occurred decides whether the error has to be considered as a warning or a real error. As a general rule, calls which free resources lead to warnings and everything else is an error. The error report of such a problem includes a stack backtrace (if enabled) and the error message generated by MPI.

To catch MPI errors this way, Intel® Trace Collector overrides any error handlers installed by the application. Errors will always be reported, even if the application or test program sets an error handler to skip over known and/or intentionally bad calls. Because the MPI standard does not guarantee that errors are detected and that proceeding after a detected error is possible, such programs are not portable and should be fixed. Intel® Trace Collector on the other hand knows that proceeding despite an error is allowed by all supported MPIs and thus none of the parameter errors is considered a hard error.

Communicator handles are checked right at the start of an MPI wrapper by calling an MPI function which is expected to check its arguments for correctness. Data type handles are tracked and then checked by Intel® Trace Collector itself. The extra parameter check is visible when investigating such an error in a debugger and although perhaps unexpected is perfectly normal. It is done to centralize the error checking.