Olivier Riff wrote:

I assume this question has been already discussed many times, but I can not find on Internet a solution to my problem.
It is about buffer size limit of MPI_Send and MPI_Recv with heterogeneous system (32 bit laptop / 64 bit cluster).
My configuration is :
open mpi 1.4, configured with: --without-openib --enable-heterogeneous --enable-mpi-threads
Program is launched a laptop (32 bit Mandriva 2008) which distributes tasks to do to a cluster of 70 processors  (64 bit RedHat Entreprise distribution):
I have to send various buffer size from few bytes till 30Mo.

You really want to get your program running without the tcp_eager_limit set if you want a better usage of memory.  I believe the crash has something to do with the rendezvous protocol in OMPI.  Have you narrowed this failure down to a simple MPI program?  Also I noticed that you're configuring with --enable-mpi-threads, have you tried configuring without that option?
I tested following commands:
1) mpirun -v -machinefile machinefile.txt MyMPIProgram
-> crash on client side ( 64 bit RedHat Entreprise ) when sent buffer size > 65536.
2) mpirun --mca btl_tcp_eager_limit 30000000 -v -machinefile machinefile.txt MyMPIProgram
-> works but has the effect of generating gigantic memory consumption on 32 bit machine side after MPI_Recv. Memory consumption goes from 800Mo to 2,1Go after receiving about 20ko from each 70 clients ( a total of about 1.4 Mo ).  This makes my program crash later because I have no more memory to allocate new structures. I read in a openmpi forum thread that setting btl_tcp_eager_limit to a huge value explains this huge memory consumption when a message sent does not have a preposted ready recv. Also after all messages have been received and there is no more traffic activity : the memory consumed remains at 2.1go... and I do not understand why.
Are the 70 clients all on different nodes?  I am curious if the 2.1GB is due to the SM BTL or possibly a leak in the TCP BTL.

What is the best way to do in order to have a working program which also has a small memory consumption (the speed performance can be lower) ?
I tried to play with mca paramters btl_tcp_sndbuf and mca btl_tcp_rcvbuf, but without success.

Thanks in advance for you help.

Best regards,


_______________________________________________ users mailing list users@open-mpi.org http://www.open-mpi.org/mailman/listinfo.cgi/users

Terry D. Dontje | Principal Software Engineer
Developer Tools Engineering | +1.650.633.7054
Oracle - Performance Technologies
95 Network Drive, Burlington, MA 01803
Email terry.dontje@oracle.com