[MPICH] MPICH2 Thread-safety

[Sylvain Jeaugey]

On Fri, 9 Jun 2006, Rajeev Thakur wrote:

> Sylvain,
>        A lot of things are happening thread-safety-wise in MPICH2. 1.0.4
> will have much better support.
Good to hear !
> Nonetheless, the paragraph you quote will
> work in 1.0.3 as well because the lock is released somewhere inside the
> progress engine before a receive blocks. (It's not obvious from looking at
> the code for MPI_Recv.)
Ok. Good.

> What exactly doesn't work for you? We have fixed a
> couple of bugs that were found.
It doesn't compile with
./configure --enable-threads=multiple:global_mutex
typing 'make' after configure gives :
In file included from ../../../../include/mpiimpl.h:83,
                  from sock.c:9:
../../../../include/mpiimplthread.h:10:49: syntax error ':' without preceding '?'
... and a lot of other errors after that.

> Another fix that is coming is that you won't need to build MPICH2 separately
> to get a thread safe version. Support for thread safety will be there by
> default and will be enabled when the user selects it with MPI_Init_thread.
> If the user doesn't call MPI_Init_thread, the code will not use any thread
> locks, so there is no penalty.
That's a very, very good thing ! I'm eager to have this new version !

Thanks for your help,

Sylvain

>> -----Original Message-----
>> From: owner-mpich-discuss at mcs.anl.gov
>> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Sylvain Jeaugey
>> Sent: Friday, June 09, 2006 7:04 AM
>> To: mpich-discuss at mcs.anl.gov
>> Subject: [MPICH] MPICH2 Thread-safety
>>
>> Hi all,
>>
>> After some tries, it looks like MPICH2 is still a little far
>> from being
>> thread-safe.
>>
>> Nothing seems to be currently done in this direction, so I'm
>> posting here
>> to know what I could do to improve this.
>>
>> The only thread policy is "global_mutex", which -in my view-
>> contradicts
>> Section 8.7.1 of the standard :
>>
>> "Example Process 0 consists of two threads. The first thread
>> executes a
>> blocking send call MPI_Send(buff1, count, type, 0, 0, comm),
>> whereas the
>> second thread executes a blocking receive call MPI_Recv(buff2, count,
>> type, 0, 0, comm, &status). I.e., the first thread sends a
>> message that is
>> received by the second thread. This communication should
>> always succeed."
>>
>> There is a lot of code for "thread-safety". Still, I can't
>> get something
>> working.
>> Is there a chance that things will be better in 1.0.4 ? Where
>> should I
>> begin if I want to improve things (and not destroy what's already be
>> done) ?
>>
>> Thanks in advance,
>>
>> Sylvain
>>
>>
>
>