[Pvfs2-users] problem with MPI-IO

[Russell L. Carter]

Hi Rob,

Robert Latham wrote:
> On Thu, May 17, 2007 at 10:28:49AM -0700, Russell L. Carter wrote:
>> I can run my single process MPI-IO program just fine
>> on a unix fs, but I get segfaults when I run it on the
>> pvfs2 fs.  I'm not getting any useful debugging output
>> when I run it under mpiexec.  Basically the app
>> opens a file, reads it, writes out the data as a new
>> file.  It repeats this for 64 files. At some point in
>> the process, seemingly random,
>> the process crashes on file open, after a significant number of
>> reads and writes, such as 33, for example.  
> 
> Hi Russell
> 
> When you see this crash, are you running with one processor or many?

One process (so far) talking to a 3 node pvfs2 fs.  I can perform
several tens of successful sessions of

open file type A, read close open file type A write close
open file type B, read close open file type B write close

before the crash in an open.  But the crash happens at
a different "session number" every time.  To make this clear,
I just ran the app 12 times in a row.  Here is the result

Failed Session #    file type  op
32                     B        open
37                     A        open
39                     B        open
40                     B        open
41                     B        open
42                     B        open
45                     A        open
45                     B        open
after a several minutes pause
50                     A        open
after a while more
52                     A        open
53                     B        open
56                     A        open

After a crash, the open leaves a zero length file with the
correct name in the pvfs2 directory, as verified by pvfs2-ls.
I open read ops with the mode MPI::MODE_RDONLY, and open
write ops with the mode MPI::MODE_WRONLY | MPI::MODE_CREATE.

1, 2, and 3 processes work correctly doing MPI-IO to a single
local xfs file.

> are you opening the file with MPI_COMM_SELF, or using a larger
> communicator?

No difference between MPI_COMM_WORLD and MPI_COMM_SELF.

Thanks,
Russell

> 
> ==rob
> 


-- 
Russell L. Carter
Esturion, LLC
2285 Sandia Drive
Prescott, Arizona 86301

rcarter at esturion.net
928 308-4154