All-to-many Communication Implemented Using MPI Asynchronous Functions

This repository contains a program designed to evaluate the performance of all-to-many personalized communication implemented using MPI asynchronous communication functions MPI_Isend, MPI_Irecv, and MPI_Waitall.

The program ./isendrecv.c runs the all-to-many communication in a loop, containing only calls to MPI_Isend, MPI_Irecv, and MPI_Waitall, but no other MPI communication functions. The number of iterations can be set by the command-line option -n. The default is 100. The message sizes are set to be different among processes, but they are all multiples of the unit message size. The default unit message size is 10 bytes, adjustable by command-line option -b.

Program ./isendrecv.c is designed to run multiple compute nodes with multiple MPI processes on each node. The command-line option -p should be set to the number of MPI processes per compute node.

When running this program on Cori KNL nodes at NERSC with 64 MPI processes on each node, we found that the timing of posting MPI_Irecv starts to become poor when the number of iterations is greater than 16.

• Compile command:

  • Edit file ./Makefile to customize the C compiler and compile options.
  • Run command make to compile and generate the executable program named isendrecv.
  • To compile for KNL nodes on Cori @NERSC, run command model swap below, before running make.

      % module swap craype-haswell craype-mic-knl
    

• Run command:

  • Example run command using mpiexec and 512 MPI processes:

      % mpiexec -n 512 ./isendrecv -p 64 -b 32 -n 20
    

  • Command-line options:

      % ./isendrecv -h
      Usage: isendrecv [OPTION]... [FILE]...
         [-h] Print help
         [-p] number of MPI processes per node
         [-b] message block unit size
         [-n] number of iterations
    

  • An example batch script file for job running on Cori @NERSC is given in ./slurm.knl.

• Example outputs on screen

  • Two output files included in this repo, ./output_cori_knl.txt and ./output_cori_haswell.txt contain the outputs of runs on Cori KNL and Haswell nodes, respectively.
  • Partial results from ./output_cori_knl.txt are given below. It shows the end-to-end timing does NOT proportionally increase as the number of iterations while all other parameters are kept constant. The end-to-end times jumps from 0.1895 seconds (8 iterations) to 0.6427 seconds (16 iterations) to 6.2541 seconds (32 iterations).

    % srun -n 512 -c 4 --cpu_bind=cores ./isendrecv -p 64 -b 32 -n 2
    -----------------------------------------------------------
    Total number of MPI processes        = 512
    Total number of sender   processes   = 512
    Total number of receiver processes   = 8
    Total number of iterations           = 2
    Message block unit size              = 32 bytes
    Max message size                     = 256 bytes
    Min message size                     = 32 bytes
    Max time of posting irecv            = 0.0117 sec
    Max time of posting isend            = 0.0041 sec
    Max time of MPI_Waitall()            = 0.0791 sec
    Max end-to-end time                  = 0.0829 sec
    (Max means the maximum among all 512 processes)
    -----------------------------------------------------------
  
    % srun -n 512 -c 4 --cpu_bind=cores ./isendrecv -p 64 -b 32 -n 8
    -----------------------------------------------------------
    Total number of MPI processes        = 512
    Total number of sender   processes   = 512
    Total number of receiver processes   = 8
    Total number of iterations           = 8
    Message block unit size              = 32 bytes
    Max message size                     = 256 bytes
    Min message size                     = 32 bytes
    Max time of posting irecv            = 0.0824 sec
    Max time of posting isend            = 0.0047 sec
    Max time of MPI_Waitall()            = 0.1161 sec
    Max end-to-end time                  = 0.1895 sec
    (Max means the maximum among all 512 processes)
    -----------------------------------------------------------
  
    % srun -n 512 -c 4 --cpu_bind=cores ./isendrecv -p 64 -b 32 -n 16
    -----------------------------------------------------------
    Total number of MPI processes        = 512
    Total number of sender   processes   = 512
    Total number of receiver processes   = 8
    Total number of iterations           = 16
    Message block unit size              = 32 bytes
    Max message size                     = 256 bytes
    Min message size                     = 32 bytes
    Max time of posting irecv            = 0.4630 sec
    Max time of posting isend            = 0.0044 sec
    Max time of MPI_Waitall()            = 0.2542 sec
    Max end-to-end time                  = 0.6427 sec
    (Max means the maximum among all 512 processes)
    -----------------------------------------------------------
  
    % srun -n 512 -c 4 --cpu_bind=cores ./isendrecv -p 64 -b 32 -n 32
    -----------------------------------------------------------
    Total number of MPI processes        = 512
    Total number of sender   processes   = 512
    Total number of receiver processes   = 8
    Total number of iterations           = 32
    Message block unit size              = 32 bytes
    Max message size                     = 256 bytes
    Min message size                     = 32 bytes
    Max time of posting irecv            = 5.6778 sec
    Max time of posting isend            = 0.0072 sec
    Max time of MPI_Waitall()            = 1.6124 sec
    Max end-to-end time                  = 6.2541 sec
    (Max means the maximum among all 512 processes)
    -----------------------------------------------------------
  

Questions/Comments:

email: wkliao@eecs.northwestern.edu

Copyright (C) 2018, Northwestern University.

See COPYRIGHT notice in top-level directory.