[mpiwg-abi] asynchronous voting to determine what requires more discussion (ie a meeting)

Christoph Niethammer niethammer at hlrs.de
Mon Apr 24 09:42:23 CDT 2023


For the io_node_list I could think of a solution by providing a new info key to give a path to a file containing a node list like the PBS/SLURM node list. However this might be tricky again if we have no shared file system on hand.

Looking forward to Quincy's expertise in the field.

Best
Christoph



----- Original Message -----
From: "Jeff Hammond" <jeff.science at gmail.com>
To: "Christoph Niethammer" <niethammer at hlrs.de>
Cc: "mpiwg-abi" <mpiwg-abi at lists.mpi-forum.org>
Sent: Monday, 24 April, 2023 16:28:03
Subject: Re: [mpiwg-abi] asynchronous voting to determine what requires more discussion (ie a meeting)

> 
> 4. Looking at the string constants I am a bit concerned about the limit of
> MPI_MAX_INFO_VAL choice of only 1024.
> We have, e.g., the Info keys "io_node_list" or "filename" where I could imagine that we could run into problems.
> The "io_node_list" key could be a problem on lager systems that come with more than 100 cabinets - assuming for hostnames typically around 9 characters/15 byte string IP + 1 comma for separation.
> The "filename" key could be a problem if we interpret the description in the standard - "This hint specifies the file name used when the file was opened." - to include the path.
> This is unlimited in length for nearly all common filesystems (thinking ext4, lustre, ...). However, Linux will limit the combined filenames/path length to 4096 characters for us.

Interesting.  I have merely worked on the assumption that MPICH and OMPI already thought about this enough and that the larger of their choices was sufficient.  It is possible that is not the case.

For filename, it seems that the problem appears only in MPI_FILE_GET_INFO, because I don’t see a limit for the input to MPI_FILE_OPEN.  I’m not sure that making the max info val larger is the right solution.  I’d be more inclined to ask the IO WG to add MPI_FILE_GET_NAME that is designed for accessing this, and design it to not have a hard-coded limit.

I wonder what Quincy thinks about io_node_list, since he may work for the company with the largest ensemble of distributed storage on the planet.

Jeff


More information about the mpiwg-abi mailing list