[MPICH] MPI_FILE_Ixxx

[Robert Latham]

On Sat, Feb 24, 2007 at 02:52:49PM +0100, David Büttner wrote:
> I am interested in using those functions on a Debian Linux Cluster and
> would like to know if anyone knows if there are plans to implement or
> already existing first (not yet published?) versions of threaded
> functions? If there are versions available, would it be possible to use
> and test those?

Hi David

There are some ideas on how to get the nonblocking I/O routines to
make asyncronous progress, but other efforts have lately taken a
higher priority. 

Would you be able to use something like MPI_THREAD_FUNNELED to carry
out MPI_File routines while other parts of your code did computation?

What file system are you planning on using?  On a linux cluster, even
with an ideal MPI_FIle_I* implementation, you won't see too much
overlap between computation and I/O with the POSIX aio calls: glibc
serializes outstanding I/O requests to the same file (we think in
order to enforce POSIX consistency semantics).  

Let's keep talking about your I/O workload.  We might be able to find
a blocking solution that gets you the performance you want.

==rob

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