Intel® Trace Analyzer and Collector User and Reference Guide

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

Detecting Illegal Buffer Modifications


MPI owns the memory that active communication references. The application must not touch it during that time. Illegal writes into buffers that the MPI is asked to send are detected by calculating a checksum of the data immediately before the request is activated and comparing it against a checksum when the send completes. If the checksum is different, someone must have modified the buffer. The reported LOCAL:MEMORY:ILLEGAL_MODIFICATION is a real error.

This problem is more common with non-blocking communication because the application gets control back while MPI still owns the buffer and then might accidentally modify the buffer. For non-blocking communication the call stacks of where the send was initiated and where it completed are provided. For persistent requests it is also shown where it was created.

The problem might also occur for blocking communication, for example when the MPI implementation incorrectly modifies the send buffer, the program is multithreaded and writes into it or other communication happens to write into the buffer. In this case only the call stack of the blocking call where the problem was detected gets printed.

Strictly speaking, reads are also illegal because the MPI standard makes no guaranteed about the content of buffers while MPI owns them. Because reads do not modify buffers, such errors are not detected. Writes are also not detected when they happen (which would make debugging a lot easier) but only later when the damage is detected.