LPP » HPC

{{>toc}}

h1. Bash tricks

h2. Get disk usage

<pre><code class="bash">

df -h

</code></pre> 

Will give you something like this:

<pre>

Sys. de fichiers       Taille Utilisé Dispo Uti% Monté sur

/dev/sde1                110G    8,1G   97G   8% /

devtmpfs                 126G       0  126G   0% /dev

tmpfs                    126G       0  126G   0% /dev/shm

tmpfs                    126G    920K  126G   1% /run

tmpfs                    126G       0  126G   0% /sys/fs/cgroup

tmpfs                    126G    4,0K  126G   1% /tmp

/dev/mapper/ddf1_DATA2    15T    7,7G   14T   1% /home

</pre>

h2. Get CPU and RAM usage

<pre><code class="bash">

htop

</code></pre> 

Will give you something like this:

p=. !{width: 80%}htop.png(htop screenshot example)!

To get more information about htop see "here":http://htop.sourceforge.net/index.php?page=main

h2. Start a Job without a queue

All process started from ssh are terminated when you close the ssh connection, even if you fork them (./app &). They are closed because when a process is closed the system send the SIGTERM signal to all its children, it’s done to avoid zombies process on a machine. To keep your task alive when you disconnect from ssh, you should use screen ("tutorial":http://www.rackaid.com/blog/linux-screen-tutorial-and-how-to/), it will block the SIGTERM signal. To start your application with screen:

<pre><code class="bash">

screen # to start screen

./your_app   #to start your application or any other command

             #type ’ Ctrl-A’ d to leave screen with your application running in background

</code></pre> 

To reconnect to your previous session:

<pre><code class="bash">

screen -ls # to list running sessions

screen -r 33287.pts-36.bender # to reconnect to 33287.pts-36.bender session

exit # to close your screen session

</code></pre>

h2. Start a mpi code:

Please note that the default mpi distribution on "Bender":https://hephaistos.lpp.polytechnique.fr/redmine/projects/hpc/wiki/Computers#Bender  /"Flexo":https://hephaistos.lpp.polytechnique.fr/redmine/projects/hpc/wiki/Computers#Flexo  /"Jakolass":https://hephaistos.lpp.polytechnique.fr/redmine/projects/hpc/wiki/Computers#Jakolass is a custom one compiled with intel compiler and libraries, to use the gcc one you should call mpi(cc/f90/run/...) with the full path /usr/lib64/openmpi/bin/mpi(cc/f90/run/...). To run a software compiled with mpi (mpic/cxx/fortran), uses mpirun with -np to set the number of mpi processes you want to start.

<pre><code class="bash">

mpirun -np 32 /path_to_myapp/myapp

#to run myapp with 32 MPI processes 

</code></pre>

h2. Start a openMP code:

When you run a software compiled with openMP library, you can tune the number of openMP threads your code will run. To do this you just have to set the OMP_NUM_THREADS environment variable.

<pre><code class="bash">

export OMP_NUM_THREADS=32

/path_to_myapp/myapp

#to run myapp with 32 openMP threads

</code></pre>

h2. Abort a Job:

To stop a running job you can use the kill command in different ways

* You have just on Job running on the machine and you know its name

<pre><code class="bash">

 killall TheNameOfTheJob

</code></pre>

* You have multiple jobs running, you have first to get the Process ID of the job

<pre><code class="bash">

ps ax # will list the running processes 

# You can filter, for example if you have started your job with mpi

ps ax | grep mpi

</code></pre>

The first number you get is the Process ID

To kill it

<pre><code class="bash">

kill -9 123456 # if 123456 is your Process ID

</code></pre>