SlurmBasics

The CentOS/Slurm nodes

We have converted Beocat from Gentoo Linux to CentOS Linux on December 26th of 2017. Any applications or libraries from the old system must be recompiled. We also converted Beocat to use the Slurm scheduler instead of SGE. You will therefore also need to convert all your old qsub scripts over to sbatch scripts. We have developed tools to make this process as easy as possible.

Using Modules

If you're using a common code that others may also be using, we may already have it compiled in a module. You can list the modules available and load an application as in the example below for Vasp.

eos> module avail
eos> module load VASP
eos> module list

When a module gets loaded, all the necessary libraries are also loaded and the paths to the libraries and executables are automatically set up. Loading Vasp for example also loads the OpenMPI library needed to run it and adds the path to the MPI commands and Vasp executables. To see how the path is set up, try executing which vasp_std. The module system allows you to easily switch between different version of applications, libraries, or languages as well.

If you are using a custom code or one that is not installed in a module, you'll need to recompile it yourself. This process is easier under CentOS as some of the work just involves loading the necessary set of modules. The first step is to decide whether to use the Intel compiler toolchain or the GNU toolchain, each of which includes the compilers and other math libraries. The module commands for each are below, and you can load these automatically when you log in by adding one of these module load statements to your .bashrc file. See /homes/daveturner/.bashrc as an example, where I put the module load statements .

To load the Intel compiler tool chain including the Intel Math Kernel Library:
eos> module load iomkl

To load the GNU compiler tool chain including OpenBLAS, FFTW, and ScalaPack load foss (free open source software):
eos> module load foss

Modules provide an easy way to set up the compilers and libraries you may need to compile your code. Beyond that there are many different ways to compile codes so you'll just need to follow the directions. If you need help you can always email us at beocat@cs.ksu.edu.

Converting your qsub script for sbatch using kstat.convert

If you already have a qsub script, I have created a new perl program called kstat.convert that will automatically convert your qsub script over to an sbatch script.

kstat.convert --sge qsub_script.sh --slurm slurm_script.sh

Below is an example of a simple qsub script and the resulting sbatch script after conversion.

#!/bin/bash
#$ -j y
#$ -cwd
#$ -N netpipe
#$ -P KSU-CIS-HPC

#$ -l mem=4G
#$ -l h_rt=100:00:00
#$ -pe single 32

#$ -M daveturner@ksu.edu
#$ -m ab

mpirun -np $NSLOTS NPmpi -o np.out

#!/bin/bash -l
#SBATCH --job-name=netpipe

#SBATCH --mem-per-cpu=4G   # Memory per core, use --mem= for memory per node
#SBATCH --time=4-04:00:00   # Use the form DD-HH:MM:SS
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=32

#SBATCH --mail-user=daveturner@ksu.edu
#SBATCH --mail-type=ALL   # same as =BEGIN,FAIL,END

mpirun -np $SLURM_NPROCS NPmpi -o np.out

The sbatch file uses #SBATCH to identify command options for the scheduler where the qsub file uses #$. Most options are similar but simply use a different syntax. The memory can still be defined on a per core basis as with SGE, or you can use --mem=128G to specify the total memory per node if you'd prefer. The --nodes= and --ntasks-per-node= provide an easy way to request the core configuration you want. If your code can be distributed across multiple nodes and you don't care what the arrangement is, you can instead just specify the number of cores using --ntasks=. For more in depth documentation on converting from SGE to Slurm follow the links below:

https://srcc.stanford.edu/sge-slurm-conversion
https://slurm.schedmd.com/sbatch.html

Submitting jobs to Slurm

Once your qsub script has been converted to an sbatch script and you have an application compiled for CentOS, you can submit the job using the sbatch command.

eos> sbatch sbatch_script.sh
eos> kstat --me

This will submit the script and show you a list of your jobs that are running and the jobs you have in the queue. By default the output for each job will go into a slurm-###.out file where ### is the job ID number. If you need to kill a job, you can use the scancel command with the job ID number.

Submitting your first job

To submit a job to run under Slurm, we use the sbatch (submit batch) command. The scheduler finds the optimum place for your job to run. With over 300 nodes and 7500 cores to schedule, as well as differing priorities, hardware, and individual resources, the scheduler's job is not trivial and it can take some time for a job to start even when there are empty nodes available.

There are a few things you'll need to know before running sbatch.

• How many cores you need. Note that unless your program is created to use multiple cores (called "threading"), asking for more cores will not speed up your job. This is a common misperception. Beocat will not magically make your program use multiple cores! For this reason the default is 1 core.
• How much time you need. Many users when beginning to use Beocat neglect to specify a time requirement. The default is one hour, and we get asked why their job died after one hour. We usually point them to the FAQ.
• How much memory you need. The default is 1 GB. If your job uses significantly more than you ask, your job will be killed off.
• Any advanced options. See the AdvancedSlurm page for these requests. For our basic examples here, we will ignore these.

So let's now create a small script to test our ability to submit jobs. Create the following file (either by copying it to Beocat or by editing a text file and we'll name it myhost.sh. Both of these methods are documented on our LinuxBasics page.

#!/bin/sh
srun hostname

Be sure to make it executable

chmod u+x myhost.sh

So, now lets submit it as a job and see what happens. Here I'm going to use five options

• --mem-per-cpu= tells how much memory I need. In my example, I'm using our system minimum of 512 MB, which is more than enough. Note that your memory request is per core, which doesn't make much difference for this example, but will as you submit more complex jobs.
• --time= tells how much runtime I need. This can be in the form of "minutes", "minutes:seconds", "hours:minutes:seconds", "days-hours", "days-hours:minutes" and "days-hours:minutes:seconds". This is a very short job, so 1 minute should be plenty. This can't be changed after the job is started please make sure you have requested a sufficient amount of time.
• --cpus-per-task=1 tells Slurm that I need only a single core per task. The AdvancedSlurm page has much more on the "cpus-per-task" switch.
• --ntasks=1 tells Slurm that I only need to run 1 task. The AdvancedSlurm page has much more on the "ntasks" switch.
• --nodes=1 tells Slurm that this must be run on one machine. The AdvancedSlurm page has much more on the "nodes" switch.
• --nodes=4 --ntasks-per-node=16 --constraint=elves requests 4 nodes with 16 cores on each and to only use the Elves.

% ls
myhost.sh
% sbatch --time=1 --mem-per-cpu=512M --cpus-per-task=1 --ntasks=1 --nodes=1 ./myhost.sh
salloc: Granted job allocation 1483446

Since this is such a small job, it is likely to be scheduled almost immediately, so a minute or so later, I now see

% ls
myhost.sh
slurm-1483446.out

% cat slurm-1483446.out
mage03

Monitoring Your Job

The kstat perl script has been developed at K-State to provide you with all the available information about your jobs on Beocat. kstat --help will give you a full description of how to use it. The Slurm version of kstat is very similar to the SGE version, with the exception that the actual memory usage of each job is not always available so the memory requested is reported, and the memory usage on each node is not always accurate since Slurm includes disk cache. We are continuing to look for better ways to get the memory usage for each job, but at the moment you may need to use Ganglia and look at the memory graph for the node you are running on to get an accurate idea of the memory being used by your application.

Eos> kstat --help

USAGE: kstat [-q] [-c] [-g] [-l] [-u user] [-p NaMD] [-j 1234567] [--part partition]
      kstat alone dumps all info except for the core summaries
      choose -q -c for only specific info on queued or core summaries.
      then specify any searchables for the user, program name, or job id

kstat                 info on running and queued jobs
kstat -q              info on the queued jobs only
kstat -c              core usage for each user
kstat -g              gpu nodes only
kstat -l -h           long list - prints full node list
kstat -u daveturner   job info for one user only
kstat --me            job info for my jobs only
kstat -j 1234567      info on a given job id
kstat --nocolor       do not use any color

--------------------------------------------------------------------------
  Multi-node jobs are highlighted in Magenta
     The switch and nodes/switch are on the right
     highlighted in Yellow when nodes are spread across multiple switches
  Shared jobs are highlighted in Cyan
  Memory requested is reported along with the total used when available
     Total RSS / Total VMSize / Total requested
  Runtime is colorized with yellow then red for jobs nearing their time limit
  Time in the queue is colorized yellow then red for jobs waiting long times
--------------------------------------------------------------------------

kstat can be used to give you a summary of your jobs that are running and in the queue:
Eos> kstat --me

Hero43       24 of 24 cores       Load 23.4 / 24       495.3 / 512 GB used
     daveturner       unafold        1234567       1 core                running                4gb req                 0 d 5 h 35 m
     daveturner       octopus       1234568      16 core               running                128gb req                 8 d 15 h 42 m
################################## BeoCat Queue ###################################
     daveturner       NetPIPE       1234569      2 core     PD   2h   4gb req       0 d 1 h 2 m

kstat produces a separate line for each host. Use kstat -h to see information on all hosts without the jobs. For the example above we are listing our jobs and the hosts they are on.

Core usage - yellow for empty, red for empty on owned nodes, cyan for partially used, blue for all cores used.
Load level - yellow or yellow background indicates the node is being inefficiently used. Red just means more threads than cores.
Memory usage - yellow or red means most memory is used.
If the node is owned the group name will be in orange on the right. Killable jobs can still be run on those nodes.

Each job line will contain the username, program name, job ID, number of cores, the status which may be colored red for killable jobs, the maximum memory used or memory requested, and the amount of time the job has run. Jobs in the queue may contain information on the requested memory and run time, priority access, constraints, and how long the job has been in the queue. In this case, I have 2 jobs running on Hero43. unafold is using 1 core while octopus is using 16 cores. Slurm did not provide any information on the actual memory use so the memory request is reported

Detailed information about a single job

kstat can provide a get a great deal of information on a particular job including a very rough estimate of when it will run. This time is a worst case scenario as this will be adapted as other jobs finish early. This is a good way to check for job submission problems before contacting us. kstat colorizes the more important information to make it easier to identify.

Eos> kstat -j 157054

##################################   Beocat Queue    ###################################
 daveturner  netpipe     157054   64 cores  PD       dwarves fabric  CS HPC     8gb req   0 d  0 h  0 m

JobId 157054  Job Name  netpipe
  UserId=daveturner GroupId=daveturner_users(2117) MCS_label=N/A
  Priority=11112 Nice=0 Account=ksu-cis-hpc QOS=normal
  Status=PENDING Reason=Resources Dependency=(null)
  Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0
  RunTime=00:00:00 TimeLimit=00:40:00 TimeMin=N/A
  SubmitTime=2018-02-02T18:18:31 EligibleTime=2018-02-02T18:18:31
  Estimated Start Time is 2018-02-03T06:17:49 EndTime=2018-02-03T06:57:49 Deadline=N/A
  PreemptTime=None SuspendTime=None SecsPreSuspend=0
  Partitions killable.q,ksu-cis-hpc.q AllocNode:Sid=eos:1761
  ReqNodeList=(null) ExcNodeList=(null)
  NodeList=(null) SchedNodeList=dwarf[01-02]
  NumNodes=2-2 NumCPUs=64 NumTasks=64 CPUs/Task=1 ReqB:S:C:T=0:0:*:*
  TRES 2 nodes 64 cores 8192  mem gres/fabric 2
  Socks/Node=* NtasksPerN:B:S:C=32:0:*:* CoreSpec=*
  MinCPUsNode=32 MinMemoryNode=4G MinTmpDiskNode=0
  Constraint=dwarves DelayBoot=00:00:00
  Gres=fabric Reservation=(null)
  OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)
  Slurm script  /homes/daveturner/perf/NetPIPE-5.x/sb.np
  WorkDir=/homes/daveturner/perf/NetPIPE-5.x
  StdErr=/homes/daveturner/perf/NetPIPE-5.x/0.o157054
  StdIn=/dev/null
  StdOut=/homes/daveturner/perf/NetPIPE-5.x/0.o157054
  Switches=1@00:05:00

#!/bin/bash -l
#SBATCH --job-name=netpipe
#SBATCH -o 0.o%j
#SBATCH --time=0:40:00
#SBATCH --mem=4G
#SBATCH --switches=1
#SBATCH --nodes=2
#SBATCH --constraint=dwarves
#SBATCH --ntasks-per-node=32
#SBATCH --gres=fabric:roce:1

host=`echo $SLURM_JOB_NODELIST | sed s/[^a-z0-9]/\ /g | cut -f 1 -d ' '`
nprocs=$SLURM_NTASKS
openmpi_hostfile.pl $SLURM_JOB_NODELIST 1 hf.$host
opts="--printhostnames --quick --pert 3"

echo "*******************************************************************"
echo "Running on $SLURM_NNODES nodes $nprocs cores on nodes $SLURM_JOB_NODELIST"
echo "*******************************************************************"

mpirun -np 2 --hostfile hf.$host NPmpi $opts -o np.${host}.mpi
mpirun -np 2 --hostfile hf.$host NPmpi $opts -o np.${host}.mpi.bi --async --bidir
mpirun -np $nprocs NPmpi $opts -o np.${host}.mpi$nprocs --async --bidir

Completed jobs and memory usage

kstat -d #

This will provide information on the jobs you have currently running and those that have completed in the last '#' days. This is currently the only reliable way to get the memory used per node for your job. This also provides information on whether the job completed normally, was canceled with scancel, timed out, or was killed because it exceeded its memory request.

Eos> kstat -d 10

###########################  sacct -u daveturner  for 10 days  ###########################
                                     max gb used on a node /   gb requested per node
 193037   ADF         dwarf43           1 n  32 c   30.46gb/100gb    05:15:34  COMPLETED
 193289   ADF         dwarf33           1 n  32 c   26.42gb/100gb    00:50:43  CANCELLED
 195171   ADF         dwarf44           1 n  32 c   56.81gb/120gb    14:43:35  COMPLETED
 209518   matlab      dwarf36           1 n   1 c    0.00gb/  4gb    00:00:02  FAILED

Summary of core usage

kstat can also provide a listing of the core usage and cores requested for each user.

Eos> kstat -c

##############################   Core usage    ###############################
  antariksh       1512 cores   %25.1 used     41528 cores queued
  bahadori         432 cores   % 7.2 used        80 cores queued
  eegoetz            0 cores   % 0.0 used         2 cores queued
  fahrialkan        24 cores   % 0.4 used        32 cores queued
  gowri             66 cores   % 1.1 used        32 cores queued
  jeffcomer        160 cores   % 2.7 used         0 cores queued
  ldcoates12        80 cores   % 1.3 used       112 cores queued
  lukesteg         464 cores   % 7.7 used         0 cores queued
  mike5454        1060 cores   %17.6 used       852 cores queued
  nilusha          344 cores   % 5.7 used         0 cores queued
  nnshan2014       136 cores   % 2.3 used         0 cores queued
  ploetz           264 cores   % 4.4 used        60 cores queued
  sadish           812 cores   %13.5 used         0 cores queued
  sandung           72 cores   % 1.2 used        56 cores queued
  zhiguang          80 cores   % 1.3 used       688 cores queued

If you want to read more, continue on to our AdvancedSlurm page.