[Pvfs2-developers] Re: mpi-io tests

[Rob Ross]

Thanks for figuring this out Sam! -- Rob

Sam Lang wrote:
> 
> Hi Guys,
> 
> I finally found the bug(s) causing these hangs.  The first problem was 
> in the request scheduler, with the new crdirent pass-through changes. 
> The handle the crdirent gets queued on is actually the directory handle 
> instead of the dirent handle, so the operation on that handle is 
> technically read-only.  Treating it as modifying was causing other 
> operations that came along (setattr, for example) to get queued instead 
> of scheduled.
> 
> That was the first cause of hangs.  The second was in the way the sync 
> coalescing code worked.  There were cases where operations were getting 
> queued as ready-to-be-synced (coalesced), but the following operations 
> that got serviced were failing (appropriately with EEXISTS), and never 
> calling any of the coalescing code.  Julian and I had talked about this 
> being a problem a while back, but I guess it never got looked at.  In 
> any case, I was able to cleanup the sync coalescing code some, so it was 
> probably worth it.
> 
> The tests of doing multiple simultaneous creates and unlinks to the same 
> file seem to work fine now, including the open test in test/posix.  Let 
> me know if any of you still have problems.
> 
> Thanks,
> 
> -sam
> 
> Murali Vilayannur wrote:
>> Hi RobL,
>>> I'm seeing this on chiba with posix ior and the flash io benchmark
>>> (oddly enough, just with the parallel netcdf version, not the hdf5
>>> one).
>>>
>>> I agree it's something related to chiba, but have no idea what it
>>> could be.  I tried a different version of berkely db and saw the same
>>> results.
>>>
>>> pvfs2-ping works, but pvfs2-ls hangs in getattr.  servers don't
>>> *appear* to be stuck in anything, but I only hooked up a debugger to
>>> the server-that-wouldn't-die.
>>
>> I really wish we could rule this as a Chiba-specific bug, since it is so
>> hard to reproduce elsewhere or we would have known by now! :)
>> Since, I was using the vfs interface and not the mpi-io interface,
>> I can only think of the simultaneous create and unlink being the issue
>> here.
>> The other possibility is a bug in the aio libraries on Chiba...
>> If any one has any insights or seen similar behavior, do let us know!
>> thanks,
>> Murali
>>
>>
> _______________________________________________
> Pvfs2-developers mailing list
> Pvfs2-developers at beowulf-underground.org
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-developers
>