View source for MPI: Using Intel MPI

Quick notes: To be updated

* Ensure ib0 is setup and working (i.e it has an IP address and you can ping other IB hosts)
* Run using the <tt>-env I_MPI_DEVICE rdma</tt>
* Setup <tt>/etc/security/limits.conf</tt>, unlimited
* TODO: add the mpdboot command for IB

== Run IMB over IB ==
<syntaxhighlight>
[root@compute-0-0 src]# /share/apps/intel/impi/4.1.0.024/intel64/bin/mpirun -np 2 -machinefile ./hostsfile -env I_MPI_DEVICE rdma ./IMB-MPI1.mpi.intel pingpong 
 benchmarks to run pingpong 
#---------------------------------------------------
#    Intel (R) MPI Benchmark Suite V3.2.3, MPI-1 part    
#---------------------------------------------------
# Date                  : Fri Nov 30 13:38:29 2012
# Machine               : x86_64
# System                : Linux
# Release               : 2.6.32-220.13.1.el6.x86_64
# Version               : #1 SMP Tue Apr 17 23:56:34 BST 2012
# MPI Version           : 2.2
# MPI Thread Environment: 

# New default behavior from Version 3.2 on:

# the number of iterations per message size is cut down 
# dynamically when a certain run time (per message size sample) 
# is expected to be exceeded. Time limit is defined by variable 
# "SECS_PER_SAMPLE" (=> IMB_settings.h) 
# or through the flag => -time 
  


# Calling sequence was: 

# ./IMB-MPI1.mpi.intel pingpong

# Minimum message length in bytes:   0
# Maximum message length in bytes:   4194304
#
# MPI_Datatype                   :   MPI_BYTE 
# MPI_Datatype for reductions    :   MPI_FLOAT
# MPI_Op                         :   MPI_SUM  
#
#

# List of Benchmarks to run:

# PingPong

#---------------------------------------------------
# Benchmarking PingPong 
# #processes = 2 
#---------------------------------------------------
       #bytes #repetitions      t[usec]   Mbytes/sec
            0         1000         1.78         0.00
            1         1000         1.83         0.52
            2         1000         1.84         1.04
            4         1000         1.83         2.09
            8         1000         1.82         4.20
           16         1000         1.82         8.38
           32         1000         1.89        16.15
           64         1000         1.93        31.66
          128         1000         2.33        52.36
          256         1000         3.37        72.50
          512         1000         3.97       122.94
         1024         1000         5.64       173.02
         2048         1000         6.59       296.58
         4096         1000         7.55       517.62
         8192         1000         8.99       868.54
        16384         1000        12.14      1286.70
        32768         1000        13.50      2314.90
        65536          640        21.91      2853.07
       131072          320        38.66      3232.98
       262144          160        72.10      3467.26
       524288           80       142.64      3505.40
      1048576           40       276.61      3615.16
      2097152           20       544.45      3673.41
      4194304           10      1082.21      3696.15
</syntaxhighlight>

== /etc/security/limits.conf ==
Note: if you have your mpd's started before updating this file, you will need to restart the mpds (<tt>mpdallexit</tt>)

<i>The Intel MPI Library does not propagate shell limits across the job. Whatever limits are valid at the time of the MPD ring startup on a particular node will be used by all instances of the Intel MPI Library jobs afterwards.</i>
<syntaxhighlight>
# ensure the following 2 lines are present
*        soft        memlock   unlimited
*        hard        memlock   unlimited
</syntaxhighlight>

== Using mpdboot ==
Start up the mpd process on all cluster nodes
<syntaxhighlight>
  # create a hosts file, 
  mpdboot -n 16 -f ./hosts -r ssh -1
  mpdboot -n 16 -f ./hosts
</syntaxhighlight>

Verify its working ok
<syntaxhighlight>
  [mpiuser@compute000 intel64]$ mpdtrace
  compute000
  compute004
  compute011
  .. 
</syntaxhighlight>