[Pvfs2-developers] help debugging request processor/distribution

[Phil Carns]

Pete Wyckoff wrote:
> slang at mcs.anl.gov wrote on Wed, 14 Jun 2006 13:09 -0500:
> 
>>I've attached a patch of the sys-io state machine from trunk that  
>>doesn't post the flow until the initial response is received.  Phil  
>>and Murali, can you let me know if this fixes your respective  
>>problems?  Also, if you get a chance Pete, can you see what kind of  
>>performance degradation this causes for you?
> 
> 
> I tested with and without the patch, at a variety of sizes with a
> single client, and could not see any performance difference.
> 
> Since it fixes Phil's issue, suggest you check it in.
> 

We did some limited performance testing and also didn't see any 
performance difference.

I did find one odd result, although it is somewhat unrelated.  In an 
earlier email on this topic I mentioned that I was able to find a 
temporary workaround for the bug by setting PVFS2_BUFMAP_DESC_COUNT to 
1, which limits pvfs2-client-core to only performing one I/O operation 
at a time.  A surprising side effect of this was that (regardless of the 
flow post order) this configuration substantially increased performance 
for a 5 process I/O benchmark.

In this particular test there are 5 processes, each both reading and 
writing large files on the same PVFS2 file system.  The results are 
counter-intuitive, because the whole point of having multiple buffers 
was to allow more concurrent activity when multiple processes perform 
I/O.  However, the test went faster when only one I/O operation was 
allowed to proceed at a time.

I don't have time to dig into this much further right now, but I just 
wanted to share on the mailing list in case anyone else was interested.

-Phil