[dcmf] [PATCH 1/1] Issue 4362: Honor the romio_ds_write and romio_ds_read hints.

Robert Latham robl at mcs.anl.gov
Fri Feb 8 16:22:17 CST 2008


On Fri, Feb 08, 2008 at 01:56:43PM -0600, Bob Cernohous wrote:
> In this case ... are we talking about modifying BGL open()?  Or a higher 
> level?   Isn't a full solution really to build ad_pvfs and select ad_pvfs 
> vs ad_bgl at some higher level (with prefixes/hints/etc or statfs)?  I 
> don't think we want ad_bgl trying to statfs() and select different file 
> system behaviors at our level.   Show us a patch and we can comment 
> better.

Thanks for the romio_ds_write addition, Bob

No worries about the PVFS comment.  Did have a few questions, though.

Do you know why ad_bgl is based off of ad_nfs and not ad_ufs?  The
aggressive locking in ad_nfs is an attempt to flush client-side
caches, but shouldn't be needed for other file systems.  

ad_pvfs2 as it exists in ROMIO won't work on BlueGene -- we've tried.
The BG compute node kernels lack a couple important system calls
needed for the PVFS client-side libraries.

PVFS has thus far silently (and incorrectly) let any file locks from
ROMIO suceed.  This is the real crux of issue 4362 -- multiple I/O
nodes think they have exclusive write locks on the same region of the
file, resulting in corrupted data.   We can bypass the locks in this
situation with a hint, but there could be others we haven't found (or
diagnosed) yet.

The proper long term fix for PVFS is to make sure we have a path
through the MPI-IO library that never attempts to lock the file.  

- Start with ADIOI_GEN_WriteStrided
- check fs type (statfs)
- check access type (noncontig/conting in mem/file)
   - noncontig in mem, noncontig in file:  naive access
   - contig in mem, noncontig in file: naive
   - (optimization) noncontig in mem, contig in file: use
     write-combining approach like that in ad_pvfs/ad_pvfs_write.c.
     buffer each noncontiguous memory region into larger contiguous
     region.
   - contig in mem, contig in file: already handled 

Thanks
==rob

-- 
Rob Latham
Mathematics and Computer Science Division    A215 0178 EA2D B059 8CDF
Argonne National Lab, IL USA                 B29D F333 664A 4280 315B



More information about the dcmf mailing list