[MPICH2-dev] question about romio lock functions

[Rajeev Thakur]

Alexey,
           Yes, I think it is ok to add that check. For independent reads
and writes, a 0-byte read/write should return success immediately at the MPI
level itself, as we check for that case. For collective reads/writes, we go
further and carry out the operation because we don't know what other
processes may be doing.
 
Rajeev
 


  _____  

From: owner-mpich2-dev at mcs.anl.gov [mailto:owner-mpich2-dev at mcs.anl.gov] On
Behalf Of Ryzhykh, Alexey
Sent: Thursday, September 08, 2005 3:47 AM
To: mpich2-dev at mcs.anl.gov
Cc: Supalov, Alexander; Voronov, German
Subject: [MPICH2-dev] question about romio lock functions



Hi,

I am working at Intel Parallel System & Applications group.

We have a problem with benchmarks for MPI I/O functions for data write
functions MPI_File_write_*  when zero length buffer is passed to them.

We see a very long delay in fcnt() function, sometimes it takes several
hours.  We used mpich2-1.02 built for ch3:sock.  The delay appears because
of following call of fcnt() in ADIO_Set_lock:

int ADIOI_Set_lock(FDTYPE fd, int cmd, int type, ADIO_Offset offset, int
whence,

             ADIO_Offset len) 

{

    int err, error_code;

    struct flock lock;

 

    /* Depending on the compiler flags and options, struct flock 

       may not be defined with types that are the same size as

       ADIO_Offsets.  */

/* FIXME: This is a temporary hack until we use flock64 where

   available. It also doesn't fix the broken Solaris header sys/types.h

   header file, which declars off_t as a UNION ! Configure tests to

   see if the off64_t is a union if large file support is requested; 

   if so, it does not select large file support.

*/

 #ifdef NEEDS_INT_CAST_WITH_FLOCK

    lock.l_type   = type;

    lock.l_start  = (int)offset;

    lock.l_whence = whence;

    lock.l_len    = (int)len;

#else

    lock.l_type   = type;

    lock.l_whence = whence;

    lock.l_start  = offset;

    lock.l_len    = len;

#endif

 

    do {

        err = fcntl(fd, cmd, &lock);

    } while (err && (errno == EINTR));

.........

 

When zero length buffer is passed to MPI_File_write_*()   the field
lock.l_len  is assigned to zero as well.

In this case the whole file is locked but there is no reason to do that.

 Linux man says:

"Specifying 0 for l_len has the special meaning: lock all bytes starting at
the location

 specified by l_whence and l_start through to the end of file, no matter
how  large  the  file

 grows."

May be there is a reason to add the check of zero len at the beginning of
ADIOI_Set_lock like below:

 

if( len == 0) return MPI_SUCCESS;

 

In this case the observed delay is not happened.

 

I went through mpich2-1.02 romio sources and did not find explicit call of
ADIOI_Set_lock with zero length.

So I assume there are no hindrances to add the check of zero length.

 

What do you think about this idea?

 

With best regards,

--

Alexey Ryzhykh

Intel Sarov,

Russia

 

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mcs.anl.gov/mailman/private/mpich2-dev/attachments/20050908/cc57c7a7/attachment.htm>