Topic:

Setup MPICH to get all child-processes killed on the slave nodes.

Author:

Reuti, reuti__at__staff.uni-marburg.de; Philipps-University of Marburg, Germany

Version:

1.3 -- 2005-02-14 Minor change for MPICH 1.2.6, comments and corrections are welcome
1.3.1 -- 2010-11-22 Final version with minor adjustments

Contents:

Note:

This HOWTO complements the information contained in the $SGE_ROOT/mpi directory of the Grid Engine distribution.



!!! Warning !!!

MPICH(1) is no longer maintained and the final version was 1.2.7p1. The followup project MPICH2 is since version 1.3 based on Hydra as default startup method for their slaves tasks and other startup methods will be removed over time. Hydra has a compiled in Tight Integration into SGE by default, and no special setup in SGE to support MPICH2 jobs is necessary any longer.

Hydra will work out-of-the-box with a defined parallel environment where start- and stop_proc_args are both set to NONE in the to be used PE (in the essence, the same PE can now be used for Open MPI and MPICH2), and in the jobscript a plain mpiexec will discover automatically the granted slots and nodes without any further options. Nevertheless, in case that there is more than one MPI installation in a cluster available, the correct mpiexec corresponding to the compiled application must be used as usual.

This document just stays here in case you come across a legacy installation and have to fix or setup such an older version.



Symptom of this behaviour

You have parallel jobs using MPICH under SGE on LINUX. Some of these jobs are killed nicely when you use qdel and don't leave any running processes on the nodes behind. Other MPICH jobs are killed, but the calculating tasks are still present after the job is killed and consuming CPU time, while these jobs don't appear any longer in SGE.



Explanation

Every MPICH task on a slave, created by an rsh-command, tries to become the process leader. If the MPICH task is just the child of qrsh_starter, this is okay as it is already the process leader and you can kill these jobs in the usual way. This is achieved, by killing the child of qrsh_starter with "kill (-pid)" and hence the whole process group will be killed. You can check this with the command:
> ps f -eo pid,uid,gid,user,pgrp,command --cols=120

If the startup script for the process on the slave nodes consists of at least two commands (like the startup with Myrinet on the nodes), a helping shell will be created,and the MPICH task will be a child of this shell with a new PID, but still in the same process group. The odd thing is now, that MPICH will discover this and enforce a new process group with this PID, to become the process leader. So the MPICH tasks jumps out of the creating process group and the intended kill of the started process group will fail, leaving the calculation tasks running.



Solutions

To solve this, there a various possibilities available, which you may chose so that it fits best to your setup of SGE.

1. Replace the helping shell with the MPICH task

In case of e.g. Myrinet, the helping shell is created by one line in mpirun.ch_gm.pl:

$cmdline = "cd $wdir ; env $varenv $apps_cmd[$i] $apps_flags[$i]";

You can prefix the final call to your program with an "exec", so that the line looks like:

$cmdline = "cd $wdir ; exec env $varenv $apps_cmd[$i] $apps_flags[$i]";

With the "exec", the started program will replace the existing shell, and so will stay to be the process leader.

2. Define MPICH_PROCESS_GROUP=no

When this environment variable is defined, the startup of the MPICH task won't create a new process group. Where it must be defines, depends on the used version of the MPICH device:

ch_p4:


must be set on the master node and the slaves

Myrinet:


has only to be defined on the slaves


So you may define it in any file, which will be sourced during a noninteractive login to the slave nodes. To be set on the master node, you have to define it in the submit script. If you want to propagate this to the slave nodes (and avoid to change any login file), you will have to edit the rsh-wrapper in the mpi directory of SGE, so that the qrsh command used there will include -V, so that the variables are available on the slaves:

echo $SGE_ROOT/bin/$ARC/qrsh -inherit -nostdin $rhost $cmd
exec $SGE_ROOT/bin/$ARC/qrsh -inherit -nostdin $rhost $cmd
else
echo $SGE_ROOT/bin/$ARC/qrsh -inherit $rhost $cmd
exec $SGE_ROOT/bin/$ARC/qrsh -inherit $rhost $cmd

should read:

echo $SGE_ROOT/bin/$ARC/qrsh -V -inherit -nostdin $rhost $cmd
exec $SGE_ROOT/bin/$ARC/qrsh -V -inherit -nostdin $rhost $cmd
else
echo $SGE_ROOT/bin/$ARC/qrsh -V -inherit $rhost $cmd
exec $SGE_ROOT/bin/$ARC/qrsh -V -inherit $rhost $cmd

3. Recompile MPICH

When you have the source of the used programs, it is also possible to disable the creation of a new process group from MPICH and remove this behaviour completely.

For MPICH 1.2.5.2 and 1.2.6:

The file where it must be done is session.c in mpich-1.2.5.2/mpid/util. Change the line:

#if defined(HAVE_SETSID) && defined(HAVE_ISATTY) && defined(SET_NEW_PGRP)

to

#if defined(HAVE_SETSID) && defined(HAVE_ISATTY) && defined(SET_NEW_PGRP) && 0

This way you can easily go back at a later point in time.

Because the routine will be linked into your final program, you have to recompile all your programs. It's not sufficient, just to install this new version in /usr/lib (or your path to mpich), unless you have used the shared libraries of MPICH. Whether any delivered binary from a vendor uses the shared version of MPICH, or has them statically linked in, you can check with the LINUX command ldd.

Before you run ./configure for MPICH, you should export the variable RSHCOMMAND=rsh, to get the desired rsh command compiled into the libraries. To create shared libraries and use them during compilation of a MPICH program, please refer to the MPICH documentation.



Number of tasks spread to the nodes

There is a difference, how many calls to qrsh will be made depending on the used version of MPICH. If you use MPICH with the ch_p4 device over ethernet, there will always be the first task started locally without usage of any qrsh call. Instead (n-1) times will qrsh only be called. Hence you can set "job_is_first_task" to true in the definition of your PE and allow only (n-1) calls to qrsh by your job.

If you are using Myrinet, it's different. In this case exactly n times the qrsh will be used. So set the "job_is_first_task" to false.



What rsh command is compiled into a binary?

If you got only the binary from a vendor, and your are in doubt what rsh command was compiled into the program, you can try to get some more information with:
> strings <programname> | awk ' /(rsh|ssh)/ { print } '

This may give you more information than you like, but you should at least get some clue about it.



Option -nolocal to mpirun

You should avoid setting this option, neither in any mpirun.args script, nor as option to the command. The first problem would be, that the head node of the MPI job (this is not the master node, but one of the selected slaves) will not be used at all, and so the processes will only be spread to the other nodes in the list (if you got just 2 slots on one machine it can't run at all this way). The second problem will be the used rsh command: because the first rsh command is just to start the head node (which would be on a slave this way), the setting of P4_RSHCOMMAND is ignored.



Uneven distribution of processes to the slave nodes with two network cards

As Andreas pointed out in http://gridengine.sunsource.net/servlets/ReadMsg?msgId=15741&listName=users, you have to check whether the first scan of the machines file by MPICH can remove an entry at all, because `hostname` may give a different name than included in the machines file (because you are using a host_aliases file). Depending on your setup of the cluster, it may be necessary to change just one entry back to the one delivered by `hostname`. If you change all entries back to the external interface again, you program may use the wrong network for the communication by MPICH. This may, or may not, be the desired result. Code to change just one entry back to the `hostname` is a small addition to the PeHostfile2MachineFile subroutine in startmpi.sh:
PeHostfile2MachineFile()
{
myhostname=`hostname`
myalias=`grep $myhostname $SGE_ROOT/default/common/host_aliases | cut -f 1 -d " "`

cat $1 | while read line; do
# echo $line
host=`echo $line|cut -f1 -d" "|cut -f1 -d"."`
nslots=`echo $line|cut -f2 -d" "`
i=1
while [ $i -le $nslots ]; do
# add here code to map regular hostnames into ATM hostnames

if [ $host = "$myalias" ] ; then # be sure to include " for the second argument
echo $myhostname
unset myalias
else
echo $host
fi

i=`expr $i + 1`
done
done
}

Don't include this, if you already changed the startmpi.sh for the handling of Turbomole, you would get two times the external name. In general, I prefer having one PE for each application. Thus copy the mpi folder in $SGE_ROOT and name it e.g. turbo, adf, linda etc. and create corresponding PEs.



Wrong interface selected for the back channel of the MPICH-tasks with the ch_p4-device

With the hints given above, you may achieve the goal of an even distribution of your MPICH tasks, but the used network interfaces for the communication of the slave processes to the master node may be the wrong one. You will notice this, when you have a look at the program calls to the slave processes. Assume 'ic...' is the internal network (which should be used) and 'node...' the external one for the NFS traffic. Then the call:
rsh ic002 -l user -x -n /home/user/a.out node001 31052 \-p4amslave \-p4yourname ic002 \-p4rmrank 1 

will use the wrong interface for the information sent back to the master. It should have the form:

rsh ic002 -l user -x -n /home/user/a.out ic001 33813 \-p4amslave \-p4yourname ic002 \-p4rmrank 1

This can be changed, when you set and export the $MPI_HOST environment variable by hand before the call to mpirun to be the name of the internal interface on the execution node, and not the (default) `hostname` for the external interface. You can also do it for all users in the mpich.args where $MPI_HOST will be set, if it was unset before the call to mpirun and change:

MPI_HOST=`hostname`

to:

MPI_HOST=`grep $(hostname) $SGE_ROOT/default/common/host_aliases | cut -f 1 -d " "`

In case that you don't use a host_aliases file at all, but still want to change to another interface, you can e.g. use a sed command like:

MPI_HOST=`hostname | sed "s/^node/ic/"`

If you decide to use this way, then take care with given "Hint for Turbomole" because you also have to take in account the internal hostname there and use it. Or you will not have to apply the change of "Uneven Distribution of processes...", if you already changed the $MPI_HOST for mpirun. This all highly depends on your setup of your cluster and desired use of your network interfaces. So this all might only be places, where you can change your setup; there is no golden rule, which will fit for all configurations. Just implement them and have a look at some test jobs, how they are distributed to the nodes and which interfaces they are using.



Hint for ADF

Some programs (like ADF parallel with MPICH) have a hard coded /usr/bin/rsh inside for their rsh command. To access any wrapper and let SGE take control over the slaves, you must set:
export P4_RSHCOMMAND=rsh

in the job script. Otherwise, always the built in /usr/bin/rsh will be used. For the usage of ssh please refer to the the appropriate HowTo for ssh, and leave the settings for MPICH to rsh to access the rsh-wrapper of SGE, which will then use ssh in the end.



Hint for Turbomole (no longer applies when HP-MPI / Platform-MPI is used with recent versions of Turbomole)

Turbomole is using the ch_p4 device of MPICH. But due to the fact, that this program will always startup one slave process more (as server task without CPU load) than parallel slots requested, you have to set "job_is_first_task" to false, to allow this qrsh call to be made. As mentioned above, MPICH will make only (n-1) calls to qrsh, but Turbomole will request (n+1) tasks, so that you end up with n again.

Because one call more to qrsh is made by TURBOMOLE than SGE is aware of (because of the behaviour of MPICH to start one task locally without qrsh, but still removing one line during the first scan of the machines file), the created machines file on the master node will have one line less than necessary. This yields to a second scan of the machines file, starting at the beginning. If the first line in the machines file is also the master node where the server task for Turbomole is running, all is in best order, because there is already the server process running which doesn't use any CPU time, and another working slave task is welcome. But in all other cases, you will have the slave processes uneven distributed to the slave nodes. Therefore just add one line to the startmpi.sh in the $SGE_ROOT/mpi directory in the already existing, to create one entry more with the name of the master node:

if [ $unique = 1 ]; then
PeHostfile2MachineFile $pe_hostfile | uniq >> $machines
else
PeHostfile2MachineFile $pe_hostfile >> $machines

#
# Append own hostname to the machines file.
#

echo `hostname` >> $machines
fi

This has the positive side effect, to deliver one entry in the machines file which can be at least removed at all during the first scan by MPICH. It will remove the line with `hostname`. If there is none, nothing is removed. This may be the case, when you have two network cards in the slaves, and `hostname` gives the name of the external interface, but the created hostlist by SGE contains only the internal names.