Newsgroups: comp.parallel.mpi
From: Rodrigo Bisbal <bisbal@ctc.com>
Subject: mpicc - nonblockin buffered send ???
Organization: CTC
Date: Tue, 09 Jul 1996 14:29:22 -0400
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <31E2A502.41C6@ctc.com>

MPI version : 1.0.12 (February 1, 1996)
Implementation : mpich
Uname -s    : IRIX extreme 5.3 08031226 IP22 mips
SGI hinv:
        1 200 MHZ IP22 Processor
        FPU: MIPS R4010 Floating Point Chip Revision: 0.0
        CPU: MIPS R4400 Processor Chip Revision: 6.0
        On-board serial ports: 2
        On-board bi-directional parallel port
        Data cache size: 16 Kbytes
        Instruction cache size: 16 Kbytes
        Secondary unified instruction/data cache size: 2 Mbytes
        Main memory size: 96 Mbytes
        EISA bus: adapter 0
        Integral Ethernet: ec0, version 1
        Integral SCSI controller 1: Version WD33C93B, revision D
        Integral SCSI controller 0: Version WD33C93B, revision D


PROBLEM:

I am pushing the limits of the buffer and message size for the 
nonblocking buffered send. I need to send a lot of data and continue
computing in an overlapping fashion.

This is what I've done:

. I create a buffer of 500K and attach it:
  buffer=malloc(500000);
  MPI_Buffer_attach(500000);

. then I create a message of 50K bytes size
  message=malloc(50000);

. the receiver process won't start receiving until after 10 seconds
  allowing to test if the sender buffer works.

. then I start sending 10 messages, it fills up up to 100K (2 messages),
even though
  I attached a 500K buffer ??
  The rest of the messages are not delivered until the receiver
  starts receiving.

. I am using the MPI_Bsend and MPI_Ibsend calls.

I have concluded that these buffered calls are not using the
user specified buffer. They must be using a default MPI buffering
scheme. 

QUESTIONS:

1. Is there a limit for the used-defined buffer size? (can it be more
that 100K?)
2. Besides attaching a buffer, what else do you need to increase the
buffer size ?
3. Is there additional flags when calling MPI_Bsend and/or MPI_Ibsend to
take full 
   advantage of the buffer ?
4. Any ideas, comments ?
5. Any experiences buffering within LAM or CHIMP? (I am using mpich)


ANY COMMENTS ARE APPRECIATED:
===================================
*  Rodrigo Bisbal                 *
*  Concurrent Technologies Corp.  *
*  bisbal@ctc.com                 *
===================================


-- 
_______________________________________________________________

     /\   Rodrigo Bisbal                email: bisbal@ctc.com
    \\ \  Unix Administrator  		phone: (814) 269-6221
   \ \\ /                               fax:   (814) 269-6870 
  / \/ / /                              __
 / /   \//\                             \  \     _ _ 
 \//\   / /    ...fly if you can.        \**\ ___\/ \ 
  / / /\ /                             X*#####*+~~\_\.......  
   / \\ \                                o/\  \
    \ \\                                    \__\
     \/    CONCURRENT TECHNOLOGIES CORP. 
________________________________________________________________
~

