Page History

...

SLURM

...

SLURM is

...

the batch job scheduling system for the

...

SLAC batch compute systems

...

Partitions

Partition is the term used by Slurm to reference queues.  Depending on your job requirements, select the appropriate partition:

anagpu:This 16 node partition is for individuals wishing to use GPU resources

. Generic documentation about SLURM can be found in this Quick Start User Guide.

S3DF Computing Facility

LCLS specific information about the S3DF facility is here:  Running at S3DF.

General S3DF documentation is here: https://s3df.slac.stanford.edu

Some hardware details about the S3DF can be found here, although note that 8 of the milano cluster node cores are reserved for the filesystem so only 120 (out of 128) can be used on each node for batch processing:  https://s3df.slac.stanford.edu/public/doc/#/batch-compute?id=clusters-amp-repos

Partitions

The partition/queue The information can be provided by the sinfo command. From psana nodes

sinfo

LCLS users typically use the "milano" queue at s3df:

>[cpo@sdfiana002 ~]$ sinfo
PARTITION AVAIL  TIMELIMIT  NODES  STATE  NODELIST 
anagpu NODELIST
roma*        up 10-00:00:0      1   comp sdfrome004
roma*        up 10-00:00:0     16  drng@ sdfrome[006-018,041-043]
roma*        up 10-00:00:0      1  down$ sdfrome003
roma*        up 10-00:00:0      1 drain$ sdfrome037
roma*         infiniteup 10-00:00:0      91  drain down* psanagpu[101-109] 
anagpu*sdfrome005
roma*        up 10-00:00:0     21    mix sdfrome[019-036,038-040]
milano       up 10-00:00:0      1  inval sdfmilan221
milano       up   infinite 10-00:00:0     14  drng@ sdfmilan[036-038,120-121,126,129,204-205,212,229-232]
milano       up 10-00:00:0      4   idle psanagpu[110-113] 

The * following the name means default partition (queue) is anagpu

Job Submission

There are 2 ways to submit a job on the cluster. The main way is by using the sbatch command to take full advantage of the computing power, and the other is to submit an interactive job. 

Sbatch

 drain$ sdfmilan[009,041,049,112]
milano       up 10-00:00:0      1  drain sdfmilan032
milano       up 10-00:00:0     12   resv sdfmilan[001-005,029-030,052,057,117-119]
milano       up 10-00:00:0    102    mix sdfmilan[006-008,010-019,021-028,031,033-035,039-040,042-048,050-051,053-056,058-072,101-111,113-116,122-125,127-128,130-131,201-203,206-211,213-220,222-228]
milano       up 10-00:00:0      1   idle sdfmilan020
ampere       up 10-00:00:0      1  drng@ sdfampere010
ampere       up 10-00:00:0      1   drng sdfampere011
ampere       up 10-00:00:0      3  drain sdfampere[005,008,023]
ampere       up 10-00:00:0     18    mix sdfampere[001-004,006-007,009,012-022]
[cpo@sdfiana002 ~]$ 

The "*" following the roma queue name indicates that it is a default queue for submission.

Job Submission

sbatch

The following is a simple submission script of a parallel psana batch job run with mpi.  It can be submitted with the command "sbatch submit.sh". The commands specified in the script file will be ran on the first available compute node that fits the resources requested.

The following is a simple submission script of a parallel psana batch job run with mpi.  It can be submitted with the command "sbatch submit.sh":

.  There are two ideas: "nodes" and "tasks per node".  A "node" is a physical computer box (with a host-name, for example) but each box/node typically has multiple-cpu-cores.  Typically the tasks-per-node parameter is set to utilize all the cores on each node.

NOTE: when running interactively the "mpirun" command takes a "-n" argument with a number of cores.  However, when running with slurm no "-n" is needed, as slurm infers it from the "--nodes" and "--ntasks-per-node" values.

> cat submit.sh 

psanagpu101:~$ more submit.sh
#!/bin/bash

#SBATCH --partition=anagpumilano
#SBATCH --ntasksnodes=42
#SBATCH --ntasks-per-node=2120
#SBATCH --output=%j.log

# -u flushes print statements which can otherwise be hidden if mpi hangs
mpirun python -u /reg/g/psdm/tutorials/examplePython/mpiDataSource.py

This script shows some additional features controllable via slurm:

> cat tst_script 
#!/bin/bash
#
#SBATCH --job-name=‘name’ # Job name for allocation
#SBATCH --output=‘filename’ # File to which STDOUT will be written, %j inserts jobid
#SBATCH --error=‘filename’ # File to which STDERR will be written, %j inserts jobid
#SBATCH --partition=anagpu # Partition/Queue to submit job
#SBATCH --gres=gpu:1080ti:1 # Number of GPUs
#SBATCH --ntask=8  # Total number of tasks
#SBATCH --ntasks-per-node=4 # Number of tasks per node
#SBATCH --mail-user='username'@slac.stanford.edu # Receive e-mail from slurm
#SBATCH --mail-type=ALL # Type of e-mail from slurm; other options are: Error, Info.
#
srun -l hostname
srun python ExampleMultipleChaperones.py


> sbatch tst_script 
Submitted batch job 187

Srun

The srun command gets control of a node to run jobs interactively.  These can be useful for data exploration and significant software development.

The following are a few examples:

>  srun -N2 -n4 hello.mpi 
Process 0 on psanagpu110 out of 1
Process 0 on psanagpu110 out of 1
Process 0 on psanagpu113 out of 1
Process 0 on psanagpu113 out of 1

Monitoring/Status

squeue

To check that jobs that exist on the system use the squeue command:

 -m mpi4py.run my_psana_script.py

One can also do this same command from the command line using the "--wrap" option for sbatch:

sbatch -p milano --nodes 2 --ntasks-per-node 3 --wrap="mpirun python -u -m mpi4py.run my_psana_script.py"

srun

In principle the slurm "srun" command can also be used to launch parallel jobs, however the current S3DF "srun" version only supports an older "pmi2" protocol, which is incompatible the mpi packages from conda that LCLS uses which use the newer "pmix" protocol.  srun should be avoided for parallel jobs at S3DF (see output of "srun --mpi=list").

Monitoring/Status

squeue

To check that jobs that exist on the system use the squeue command:

[cpo@sdfiana002 ~]$ squeue -u ytl
             JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
          30703603 ampere,ro      out      ytl PD       0:00      1 (launch failed requeued held)
          30703602 ampere,ro      out      ytl PD       0:00      1 (launch failed requeued held)
          30701730 ampere,ro      out      ytl PD       0:00      1 (launch failed requeued held)
          30700739 ampere,ro      out      ytl PD       0:00      1 (launch failed requeued held)
          30700738 ampere,ro      out      ytl PD       0:00      1 (launch failed requeued held)
          30699545 ampere,ro      out      ytl PD       0:00      1 (launch failed requeued held

> squeue 
       JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
         187  30704838    milano      anagpuout ExampleM    omarq ytl RCG       04:0407      1 psanagpu110 
sdfmilan221
[cpo@sdfiana002 ~]$ 

The ST (job state) field shows that jobid 187 is currently running (Rmost jobs are pending (PD) and one is completing (CG).

sacct

Get information about status of finished jobs

>[cpo@sdfiana002 ~]$ sacct     # show all finished job of user
 -u ytl
JobID         JobID    JobName  Partition    Account  AllocCPUS      State ExitCode 
  ------------ ---------- ---------- ---------- ---------- ---------- -------- 
30682524            out     milano shared:de+ 141       112  PREEMPTED      sleep 0:0 
30682524.ba+    anagpu  batch            shared:de+       1 112 COMPLETED CANCELLED     0:015 
30682524.ex+  142   extern            shared:de+        112  ls COMPLETED      0:0 
30682525    anagpu        out     milano shared:de+        1112  COMPLETEDPREEMPTED      0:0 
30682525.ba+  143    batch            exitshared:de+     anagpu   112  CANCELLED     0:15 
30682525.ex+     extern     1     FAILED  shared:de+    2:0 

> sacct -j <jobid>112  COMPLETED     # show info for specific job  

0:0 

Misc Slurm commands

scontrol is used to view or modify Slurm configuration including: job, job step, node, partition, reservation, and overall system configuration. Most of the commands can only be executed by user root or an Administrator.

• Detail job information: scontrol show jobid -dd <jobID>
• Show reservation: scontrol show res

sacctmgr is used to deal with accounts, associations and users.
Format can be modified at will or removed to see all (can be hard to read, especially on smaller windows):

• Show what account a user is associated with: sacctmgr show associations Users=<user_names> format=cluster,account,partition,QOS
• Show priorities for an account: sacctmgr list associations -p accounts=<accounts>
• Show priority coefficients: sacctmgr show qos format=name,priority,usagefactor

Others

• Show priority level for a job: sprio -j <jobID>

...