| Version 64 (modified by , 2 months ago) (diff) |
|---|
Table of Contents
Information about the batch system (SLURM)
The DEEP prototype system is running SLURM for resource management. Documentation of Slurm can be found here.
Overview
Slurm offers interactive and batch jobs (scripts submitted into the system). The relevant commands are srun and sbatch. The srun command can be used to spawn processes (please do not use mpiexec), both from the frontend and from within a batch script. You can also get a shell on a node to work locally there (e.g. to compile your application natively for a special platform or module).
Available Partitions
Please note that there is no default partition configured. In order to run a job, you have to specify one of the following partitions, using the --partition=... switch:
Name Description dp-cn dp-cn[01-50], DEEP Cluster nodes (Xeon Skylake) dp-dam dp-dam[01-16], DEEP DAM nodes (Xeon Cascadelake + 1 V100 + 1 Stratix 10) dp-esb dp-esb[log:@26-75 "[01-75]"], DEEP ESB nodes connected with IB EDR (Xeon Cascadelake + 1 V100) dp-sdv-esb dp-sdv-esb[01-02], DEEP ESB Test nodes (Xeon Cascadelake + 1 V100) ml-gpu ml-gpu[01-03], GPU test nodes for ML applications (4 V100 cards) knl knl[01,04-06], KNL nodes knl256 knl[01,05], KNL nodes with 64 cores knl272 knl[04,06], KNL nodes with 68 cores snc4 knl[05], KNL node in snc4 memory mode debug all compute nodes (no gateways)
Anytime, you can list the state of the partitions with the sinfo command. The properties of a partition (.e.g. the maximum walltime) can be seen using
scontrol show partition <partition>
Remark about environment
By default, Slurm passes the environment from your job submission session directly to the execution environment. Please be aware of this when running jobs with srun or when submitting scripts with sbatch. This behavior can be controlled via the --export option. Please refer to the Slurm documentation to get more information about this.
In particular, when submitting job scripts, it is recommended to load the necessary modules within the script and submit the script from a clean environment.
An introductory example
Suppose you have an mpi executable named hello_mpi. There are three ways to start the binary.
From a shell on a node
If you just need one node to run your interactive session on you can simply use the srun command (without salloc), e.g.:
[kreutz1@deepv ~]$ srun -A deep -N 1 -n 8 -p dp-cn -t 00:30:00 --pty --interactive bash [kreutz1@dp-cn22 ~]$ srun -n 8 hostname dp-cn22 dp-cn22 dp-cn22 dp-cn22 dp-cn22 dp-cn22 dp-cn22 dp-cn22
The environment is transported to the remote shell, no .profile, .bashrc, … are sourced (especially not the modules default from /etc/profile.d/modules.sh). As of March 2020, an account has to be specified using the --account (short -A) option, which is "deepsea" for DEEP-SEA project members. For people not included in the DEEP-SEA project, please use the "Budget" name you received along with your account creation.
Assume you would like to run an MPI task on 4 cluster nodes with 2 tasks per node. It's necessary to use salloc then:
[kreutz1@deepv Temp]$ salloc -A deep -p dp-cn -N 4 -n 8 -t 00:30:00 srun --pty --interactive /bin/bash [kreutz1@dp-cn01 Temp]$ srun -N 4 -n 8 ./MPI_HelloWorld Hello World from rank 3 of 8 on dp-cn02 Hello World from rank 7 of 8 on dp-cn04 Hello World from rank 2 of 8 on dp-cn02 Hello World from rank 6 of 8 on dp-cn04 Hello World from rank 0 of 8 on dp-cn01 Hello World from rank 4 of 8 on dp-cn03 Hello World from rank 1 of 8 on dp-cn01 Hello World from rank 5 of 8 on dp-cn03
Once you get to the compute node, start your application using srun. Note that the number of tasks used is the same as specified in the initial srun command above (4 nodes with two tasks each). It's also possible to use less nodes in the srun command. So the following command would work as well:
[kreutz1@dp-cn01 Temp]$ srun -N 1 -n 1 ./MPI_HelloWorld Hello World from rank 0 of 1 on dp-cn01
Running directly from the front ends
You can run the application directly from the frontend, bypassing the shell. Do not forget to set the correct environment for running your executable on the login node as this will be used for execution with srun.
[kreutz1@deepv Temp]$ ml GCC/10.3.0 ParaStationMPI/5.4.9-1 [kreutz1@deepv Temp]$ srun -A deep -p dp-cn -N 4 -n 8 -t 00:30:00 ./MPI_HelloWorld Hello World from rank 7 of 8 on dp-cn04 Hello World from rank 3 of 8 on dp-cn02 Hello World from rank 6 of 8 on dp-cn04 Hello World from rank 2 of 8 on dp-cn02 Hello World from rank 4 of 8 on dp-cn03 Hello World from rank 0 of 8 on dp-cn01 Hello World from rank 1 of 8 on dp-cn01 Hello World from rank 5 of 8 on dp-cn03
It can be useful to create an allocation which can be used for several runs of your job:
[kreutz1@deepv Temp]$ salloc -A deep -p dp-cn -N 4 -n 8 -t 00:30:00 salloc: Granted job allocation 69263 [kreutz1@deepv Temp]$ srun ./MPI_HelloWorld Hello World from rank 7 of 8 on dp-cn04 Hello World from rank 3 of 8 on dp-cn02 Hello World from rank 6 of 8 on dp-cn04 Hello World from rank 2 of 8 on dp-cn02 Hello World from rank 5 of 8 on dp-cn03 Hello World from rank 1 of 8 on dp-cn01 Hello World from rank 4 of 8 on dp-cn03 Hello World from rank 0 of 8 on dp-cn01 ... # several more runs ... [kreutz1@deepv Temp]$ exit exit salloc: Relinquishing job allocation 69263
Note that in this case the -N and -n options for the srun command can be skipped (they default to the corresponding options given to salloc).
Batch script
As stated above, it is recommended to load the necessary modules within the script and submit the script from a clean environment.
The following script hello_cluster.sh will unload all modules and load the modules required for executing the given binary:
#!/bin/bash #SBATCH --partition=dp-esb #SBATCH -A deep #SBATCH -N 4 #SBATCH -n 8 #SBATCH -o /p/project/cdeep/kreutz1/hello_cluster-%j.out #SBATCH -e /p/project/cdeep/kreutz1/hello_cluster-%j.err #SBATCH --time=00:10:00 ml purge ml GCC ParaStationMPI srun ./MPI_HelloWorld
This script requests 4 nodes of the ESB module with 8 tasks, specifies the stdout and stderr files, and asks for 10 minutes of walltime. You can submit the job script as follows:
[kreutz1@deepv Temp]$ sbatch hello_cluster.sh Submitted batch job 69264
… and check what it's doing:
[kreutz1@deepv Temp]$ squeue -u $USER
JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)
69264 dp-cn hello_cl kreutz1 CG 0:04 4 dp-cn[01-04]
Once finished, you can check the result (and the error file if needed)
[kreutz1@deepv Temp]$ cat /p/project/cdeep/kreutz1/hello_cluster-69264.out Hello World from rank 7 of 8 on dp-esb37 Hello World from rank 3 of 8 on dp-esb35 Hello World from rank 5 of 8 on dp-esb36 Hello World from rank 1 of 8 on dp-esb34 Hello World from rank 6 of 8 on dp-esb37 Hello World from rank 2 of 8 on dp-esb35 Hello World from rank 4 of 8 on dp-esb36 Hello World from rank 0 of 8 on dp-esb34
Information on past jobs and accounting
The sacct command can be used to enquire the Slurm database about a past job.
[kreutz1@deepv Temp]$ sacct -j 69268
JobID JobName Partition Account AllocCPUS State ExitCode
------------ ---------- ---------- ---------- ---------- ---------- --------
69268+0 bash dp-cn deepest-a+ 96 COMPLETED 0:0
69268+0.0 MPI_Hello+ deepest-a+ 2 COMPLETED 0:0
69268+1 bash dp-dam deepest-a+ 384 COMPLETED 0:0
On the Cluster (CM) nodes it's possible to query the consumed energy for a certain job:
[kreutz1@deepv kreutz1]$ sacct -o ConsumedEnergy,JobName,JobID,CPUTime,AllocNodes -j 69326
ConsumedEnergy JobName JobID CPUTime AllocNodes
-------------- ---------- ------------ ---------- ----------
496.70K hpl_MKL_O+ 69326 16:28:48 1
0 batch 69326.batch 16:28:48 1
496.70K xlinpack_+ 69326.0 08:10:24 1
This feature will also be for the ESB nodes.
Advanced topics
For further details on the batchsystem and psslurm which is used on the DEEP system as well as on the JSC production systems, please refer to the in-depth description for using the
Batchsystem on Jureca. Among extended examples for allocation
of nodes you can find information on job steps, dependency chains and multithreading there. If you are interested in pinning of threads and tasks to certain CPUs or cores,
please also take a look into the Processor Anffinity sections of the Jureca documentation. Most of the information
provided there will also refer to the DEEP system.
FAQ
Is there a cheat sheet for all main Slurm commands?
Yes, it is available here.
Why's my job not running?
You can check the state of your job with
scontrol show job <job id>
In the output, look for the Reason field.
You can check the existing reservations using
scontrol show res
How can I check which jobs are running in the machine?
Please use the squeue command ( the "-u $USER" option to only list jobs belonging to your user id).
A graphical overview can be displayed using slurmtop command.
How do I do chain jobs with dependencies?
Please confer the sbatch/srun man page, especially the
-d, --dependency=<dependency_list>
entry.
Also, jobs can be chained after they have been submitted using the scontrol command by updating their Dependency field.
How can check the status of partitions and nodes?
The main command to use is sinfo. By default, when called alone, sinfo will list the available partitions and the number of nodes in each partition in a given status. For example:
[deamicis1@deepv hybridhello]$ sinfo PARTITION AVAIL TIMELIMIT NODES STATE NODELIST sdv up 20:00:00 11 idle deeper-sdv[06-16] knl up 20:00:00 1 drain knl01 knl up 20:00:00 3 idle knl[04-06] knl256 up 20:00:00 1 drain knl01 knl256 up 20:00:00 1 idle knl05 knl272 up 20:00:00 2 idle knl[04,06] snc4 up 20:00:00 1 idle knl05 extoll up 20:00:00 11 idle deeper-sdv[06-16] ml-gpu up 20:00:00 3 idle ml-gpu[01-03] dp-cn up 20:00:00 1 drain dp-cn33 dp-cn up 20:00:00 5 resv dp-cn[09-10,25,49-50] dp-cn up 20:00:00 44 idle dp-cn[01-08,11-24,26-32,34-48] dp-dam up 20:00:00 1 drain* dp-dam08 dp-dam up 20:00:00 2 drain dp-dam[03,07] dp-dam up 20:00:00 3 resv dp-dam[05,09-10] dp-dam up 20:00:00 2 alloc dp-dam[01,04] dp-dam up 20:00:00 8 idle dp-dam[02,06,11-16] dp-dam-ext up 20:00:00 2 resv dp-dam[09-10] dp-dam-ext up 20:00:00 6 idle dp-dam[11-16] dp-esb up 20:00:00 51 drain* dp-esb[11,26-75] dp-esb up 20:00:00 2 drain dp-esb[08,23] dp-esb up 20:00:00 2 alloc dp-esb[09-10] dp-esb up 20:00:00 20 idle dp-esb[01-07,12-22,24-25] dp-sdv-esb up 20:00:00 2 resv dp-sdv-esb[01-02] psgw-cluster up 20:00:00 1 idle nfgw01 psgw-booster up 20:00:00 1 idle nfgw02 debug up 20:00:00 1 drain* dp-dam08 debug up 20:00:00 4 drain dp-cn33,dp-dam[03,07],knl01 debug up 20:00:00 10 resv dp-cn[09-10,25,49-50],dp-dam[05,09-10],dp-sdv-esb[01-02] debug up 20:00:00 2 alloc dp-dam[01,04] debug up 20:00:00 69 idle deeper-sdv[06-16],dp-cn[01-08,11-24,26-32,34-48],dp-dam[02,06,11-16],knl[04-06],ml-gpu[01-03]
Please refer to the man page for sinfo for more information.
One useful command to check on the nodes is: sinfo -R -h -o "%n %12U %19H %6t %E" | sort -u
Can I join stderr and stdout like it was done with -joe in Torque?
Not directly. In your batch script, redirect stdout and stderr to the same file:
#!sh ... #SBATCH -o /point/to/the/common/logfile-%j.log #SBATCH -e /point/to/the/common/logfile-%j.log ...
(The %j will place the job id in the output file). N.B. It might be more efficient to redirect the output of your script's commands to a dedicated file.
