bug: shared file pointer implementation should be robustified

Open mpichbot opened this issue 9 years ago • 1 comments

Originally by robl on 2014-12-19 16:02:18 -0600

Paul Coffman reported a race condition where rank 0 might open the shared file pointer, update it, close the file, delete, and unlink it all in the time that rank 1 is trying to open the file.

The right way to fix this is to make the implementation of shared file pointers retry/reopen if the file does not exist.

The shortest fix is to barrier in MPI_FILE_CLOSE. yuck.

Oct 14 '16 18:10 mpichbot

Shouldn't MPI_File_open be a collective and thus okay for a barrier?

Aug 15 '22 00:08 hzhou