Use cases
- OpenMPI + affinity
- NUMA topology optimization
Users tips

Use cases

OpenMPI + affinity

We saw that the Linux kernel seems to be incapable of using correctly all the CPUs from the cpusets.

Indeed, reserving 2 out of 8 cores on a node and running a code that uses 2 process, these 2 process where not well assigned to each cpu. We had to give the CPU MAP to OpenMPI to do cpu_affinity:

 i=0 ; oarprint core -P host,cpuset -F "% slot=%" | while read line ; do echo "rank $i=$line"; ((i++)); done > affinity.txt
 
 [user@node12 tmp]$ mpirun -np 8 --mca btl openib,self -v -display-allocation -display-map  --machinefile $OAR_NODEFILE -rf affinity.txt /home/user/espresso-4.0.4/PW/pw.x < BeO_100.inp

NUMA topology optimization

In this use case, we've got a numa host (an Altix 450) having a “squared” topology: nodes are interconnected by routers like in this view:

In yellow, “routers”, in magenta, “nodes” (2 dual-core processors per node)

Routers interconnect IRUS (chassis) on which the nodes are plugged (4 or 5 nodes per IRU).

What we want is that for jobs that can enter into 2 IRUS or less, minimize the distance between the resources (ie use IRUS that have only one router interconnexion between them). The topology may be siplified as follows:

The idea is to use moldable jobs and an admission rule that converts automatically the user requests to a moldable job. This job uses 2 resource properties: numa_x and numa_y that may be analogue to the square coordinates. What we want in fact, is the job that ends the soonest between a job running on an X or on a Y coordinate (we only want vertical or horizontal placed jobs).

The numa_x and numa_y properties are set up this way (pnode is a property corresponding to physical nodes):

{|border=“1” !pnode !iru !numa_x !numa_y

For example, the following requested ressources:

 -l /core=16

will result into:

 -l /numa_x=1/pnode=4/cpu=2/core=2 -l /numa_y=1/pnode=4/cpu=2/core=2

Here is the admission rule making that optimization:

 # Title : Numa optimization
 # Description : Creates a moldable job to take into account the "squared" topology of an Altix 450
 my $n_core_per_cpus=2;
 my $n_cpu_per_pnode=2;
 if (grep(/^itanium$/, @{$type_list}) && (grep(/^manual$/, @{$type_list}) == "") && $#$ref_resource_list == 0){
   print "[ADMINSSION RULE] Optimizing for numa architecture (use \\"-t manual\\" to disable)";
   my $resources_def=$ref_resource_list->[0];
   my $core=0;
   my $cpu=0;
   my $pnode=0;
   foreach my $r (@{$resources_def->[0]}) {
     foreach my $resource (@{$r->{resources}}) {
       if ($resource->{resource} eq "core") {$core=$resource->{value};}
       if ($resource->{resource} eq "cpu") {$cpu=$resource->{value};}
       if ($resource->{resource} eq "pnode") {$pnode=$resource->{value};}
     }
   }
   # Now, calculate the number of total cores
   my $n_cores=0;
   if ($pnode == 0 && $cpu != 0 && $core == 0) {
     $n_cores = $cpu*$n_core_per_cpus;
   }
   elsif ($pnode != 0 && $cpu == 0 && $core == 0) {
     $n_cores = $pnode*$n_cpu_per_pnode*$n_core_per_cpus;
   }
   elsif ($pnode != 0 && $cpu == 0 && $core != 0) {
     $n_cores = $pnode*$core;
   }
   elsif ($pnode == 0 && $cpu != 0 && $core != 0) {
     $n_cores = $cpu*$core;
   }
   elsif ($pnode == 0 && $cpu == 0 && $core != 0) {
     $n_cores = $core;
   }
   else { $n_cores = $pnode*$cpu*$core; }
   print "[ADMINSSION RULE] You requested $n_cores cores\";
   if ($n_cores > 32) {
     print "[ADMISSION RULE] Big job (>32 cores), no optimization is possible";
   }else{
     print "[ADMISSION RULE] Optimization produces: /numa_x=1/$pnode/$cpu/$core
                                         /numa_y=1/$pnode/$cpu/$core";
 
     my @newarray=eval(Dumper(@{$ref_resource_list}->[0]));
     push (@{$ref_resource_list},@newarray);
     unshift(@{%{@{@{@{$ref_resource_list}->[0]}->[0]}->[0]}->{resources}},{'resource' => 'numa_x','value' => '1'});
     unshift(@{%{@{@{@{$ref_resource_list}->[1]}->[0]}->[0]}->{resources}},{'resource' => 'numa_y','value' => '1'});
   }
 }

Users tips

oarsh completion

Tip based on an idea from Jerome Reybert

In order to complete nodes names in a oarsh command, add these lines in your .bashrc

function _oarsh_complete_() {
  if [ -n "$OAR_NODEFILE" -a "$COMP_CWORD" -eq 1 ]; then
    local word=${comp_words[comp_cword]}
    local list=$(cat $OAR_NODEFILE | uniq | tr '\n' ' ')
    COMPREPLY=($(compgen -W "$list" -- "${word}"))
  fi
}
complete -o default -F _oarsh_complete_ oarsh

Then try oarsh <TAB>

OAR aware shell prompt for Interactive jobs

If you want to have a bash prompt with your job id and the remaining walltime then you can add in your ~/.bashrc:

if [ "$PS1" ]; then
    __oar_ps1_remaining_time(){
        if [ -n "$OAR_JOB_WALLTIME_SECONDS" -a -n "$OAR_NODE_FILE" -a -r "$OAR_NODE_FILE" ]; then
            DATE_NOW=$(date +%s)
            DATE_JOB_START=$(stat -c %Y $OAR_NODE_FILE)
            DATE_TMP=$OAR_JOB_WALLTIME_SECONDS
            ((DATE_TMP = (DATE_TMP - DATE_NOW + DATE_JOB_START) / 60))
            echo -n "$DATE_TMP"
        fi
    }
    PS1='[\u@\h|\W]$([ -n "$OAR_NODE_FILE" ] && echo -n "(\[\e[1;32m\]$OAR_JOB_ID\[\e[0m\]-->\[\e[1;34m\]$(__oar_ps1_remaining_time)mn\[\e[0m\])")\$ '
    if [ -n "$OAR_NODE_FILE" ]; then
        echo "[OAR] OAR_JOB_ID=$OAR_JOB_ID"
        echo "[OAR] Your nodes are:"
        sort $OAR_NODE_FILE | uniq -c | awk '{printf("      %s*%d", $2, $1)}END{printf("\n")}' | sed -e 's/,$//'
    fi
fi

Then the prompt inside an Interactive job will be like:

  [capitn@node006~](3101-->29mn)$

Many small jobs grouping

Many small jobs of a few seconds may be painful for the OAR system. OAR may spend more time scheduling, allocating and launching than the actual computation time for each job.

Gabriel Moreau developed a script that may be useful when you have a large set of small jobs. It groups you jobs into a unique bigger OAR job:

http://servforge.legi.grenoble-inp.fr/pub/soft-trokata/oarutils/oar-parexec.html

You can download it from this page:

http://servforge.legi.grenoble-inp.fr/projects/soft-trokata/wiki/SoftWare/OarUtils

For a more generic approach you can use Cigri, a grid middleware running onto OAR cluster(s) that is able to automatically group parametric jobs. Cigri is currently in a re-writing process and a new public release is planned for the end of 2012.

Please contact Bruno.Bzeznik@imag.fr for more informations.

Overcoming quoting issues n oarsub commands

One may get a bit of a weird behavior when trying to submit a series of jobs through a bash script. For instance:

#!/bin/bash
 
for d in ../test/*/ ; do
    CMD="oarsub -p \"cputype='xeon'\" -l /nodes=1,walltime=04:00:00 -S \"./myscript.sh $d/in $d/out\""
    echo $CMD
    $CMD
done

This does not work, due to quoting issues.

A solution is to use a bash arrays to define the command, as follows:

#!/bin/bash
 
for d in ../test/*/ ; do
    declare -a CMD
    CMD=(oarsub -p "cputype='xeon'" -l /nodes=1,walltime=04:00:00 -S "./myscript.sh $d/in $d/out")
    "${CMD[@]}"
done

This works.

Table of Contents