Determine MPI Data Types in col_on_comm() & dst_on_comm() to prevent displacements overflow. (Fix for #2156) #2157
Conversation
…displacements overflow. The MPI_Gatherv() and MPI_Scatterv() operations require integer displacements into the communications buffers. Historically everything is passed as an MPI_CHAR, causing these displacements to be larger than otherwise necessary. For large domain sizes this can cause the displace[] offsets to exceed the maximum int, wrapping to negative values. This change introduces additional error checking and then uses the function MPI_Type_match_size() (available since MPI-2.0) to determine a suitable MPI_Datatype given the input *typesize. The result then is that the displace[] offsets are in terms of data type extents, rather than bytes, and less likely to overflow.
|
Just for awareness, I can't see the output of the failed |
|
Common problem - it will be provided by message later
…On Fri, Jan 17, 2025 at 3:21 PM Benjamin S. Kirk ***@***.***> wrote:
Just for awareness, I can't see the output of the failed WRF-BUILD-2690;
I get a timeout accessing
https://ncar_jenkins.scalacomputing.com/job/WRF-Feature-Regression-Test/2690/console
—
Reply to this email directly, view it on GitHub
<#2157 (comment)>, or
unsubscribe
<https://github.com/notifications/unsubscribe-auth/AEIZ77CVDAXTOXJBOYNIOZ32LF67RAVCNFSM6AAAAABVMWK7MSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKOJZGMYDONRXGA>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
Fixes runtime failures caught by CI in in the underlying MPI_Gatherv().
Of course dtype needs to be MPI_Datatype, not an int. This error sneaked through MPICH-based tests but not OpenMPI. Hopefully this change will address previous CI failures.
|
The regression test results: |
|
@benkirk Thanks for the fix! I tested it for a few cases we could run before and it is working now. |
|
Thanks for the success report @weiwangncar, happy to help! |
There was a problem hiding this comment.
@benkirk Thanks for the fix! This has been a heck of a problem to hunt down.
If my understanding is correct the overflow can still manifest, just with a domain approximately 4x in size than what triggered it before, right? Though the error will now be caught and aborted at the appropriate location.
|
That's right, the (There's some potential fallbacks even in that case I can think of but haven't implemented. Those fallbacks would require some conditions on the buffers to be sent, like evenly divisible by some integer factor. Such tricks are necessary when trying to send buffers >2GB with MPI.) |
Determine MPI Data Types in col_on_comm() & dst_on_comm() to prevent displacements overflow.
TYPE: bug fix
KEYWORDS: prevent displacements overflow in MPI_Gatherv() and MPI_Scatterv() operations
SOURCE: Benjamin Kirk & Negin Sobhani (NSF NCAR / CISL)
DESCRIPTION OF CHANGES:
Problem:
The MPI_Gatherv() and MPI_Scatterv() operations require integer displacements into the communications buffers. Historically everything is passed as an MPI_CHAR, causing these displacements to be larger than otherwise necessary. For large domain sizes this can cause the displace[] offsets to exceed the maximum int, wrapping to negative values.
Solution:
This change introduces additional error checking and then uses the function MPI_Type_match_size() (available since MPI-2.0) to determine a suitable MPI_Datatype given the input *typesize. The result then is that the displace[] offsets are in terms of data type extents, rather than bytes, and less likely to overflow.
ISSUE: Fixes #2156
LIST OF MODIFIED FILES:
M frame/collect_on_comm.c
TESTS CONDUCTED:
Failed cases run now.
RELEASE NOTE:
Determine MPI Data Types in col_on_comm() & dst_on_comm() to prevent displacements overflow.