Differences

This shows you the differences between two versions of the page.

--- slurm [2021/01/06 16:00] – created kauffman
+++ slurm [2022/10/07 15:13] (current) – borja
@@ Line 8: / Line 8: @@
 Slurm is similar to most other queue systems in that you write a batch script, then submit it to the queue manager. The queue manager schedules your job to run on the queue (or partition in Slurm parlance) that you designate. Below is an outline of how to submit jobs to Slurm, how Slurm decides when to schedule your job, and how to monitor progress.
+===== Communication =====
+==== Mailing List ====
+If you are going to be a user of any of the CS Slurm clusters please sign up for the mailing list. Downtime and other relevant information will be announced here.
+[[ https://mailman.cs.uchicago.edu/cgi-bin/mailman/listinfo/slurm | Mailing List ]]
+==== Discord ====
+There is a dedicated text channel ''%%#slurm%%'' on the UChicago CS Discord server. Please note that this Discord server is //only// for UChicago-affiliated users. You can find a link to our Discord server on the [[start|front page]] of this wiki.
 ===== Clusters =====
+We have a couple different clusters. If you don't know where to start please use the ''%%Peanut%%'' cluster. The ''%%AI Cluster%%'' is for GPU jobs and more advanced users.
+  * [[slurm:peanut|Peanut Cluster]]
+  * [[slurm:ai|AI Cluster]]
 ==== Peanut Cluster ====
@@ Line 20: / Line 35: @@
 To use this cluster there are specific nodes you need to log into. Please visit the dedicated AI cluster page for more information.
@@ Line 31: / Line 43: @@
   ssh user@linux.cs.uchicago.edu
-If you want to use the AI Cluster you will need to login into:
+If you want to use the AI Cluster you will need to have previously requested access by sending in a ticket. Afterwards, you may login into:
   ssh user@fe.ai.cs.uchicago.edu
@@ Line 37: / Line 49: @@
 Please read up on the specifics on the cluster you are interested in.
-===== Mailing List =====
-If you are going to be a user of this cluster please sign up for the mailing list. Downtime and other relevant information will be announced here.
-[[ https://mailman.cs.uchicago.edu/cgi-bin/mailman/listinfo/slurm | Mailing List ]]
 ===== Documentation =====
@@ Line 57: / Line 66: @@
   * [[http://research.computing.yale.edu/support/hpc/user-guide/slurm| Yale's User Guide]]
+====== Job Submission ======
+Jobs submitted to the cluster are run from the command line. Almost anything that you can run via the command line on any of our machines in our labs can be run on our job submission server agents.
+The job submission servers run the same software as you will find on our lab computers, but without the X environment.
+You can submit jobs from the departmental computers that you have access to. You will not be able to access the job server agent directly.
+===== Command Summary =====
+[[http://slurm.schedmd.com/pdfs/summary.pdf|Cheat Sheet]]
+| ^ Slurm ^ Example ^
+^ Submit a batch serial job | sbatch | sbatch runscript.sh |
+^ Run a script interactively | srun | srun --pty -p interact -t 10 --mem 1000 \\ /bin/bash \\ /bin/hostname |
+^ Kill a job | scancel | scancel 4585 |
+^ View status of queues | squeue | squeue -u cnetid |
+^ Check current job by id | sacct | sacct -j 999999 |
+===== Usage =====
+Below are some common examples. You should consult the [[http://slurm.schedmd.com/documentation.html|documentation]] of Slurm if you need further assistance.
+=== Default Quotas ===
+By default we set a job to be run on one CPU and allocate 100MB of RAM. If you require more than that you should specify what you need. Using the following options will do: ''%%--mem-per-cpu%%'', ''%%--nodes%%'', ''%%--ntasks%%''.
+=== Exclusive access to a node ===
+You will need to add the ''%%--exclusive%%'' options to your script or command line options. This option will ensure that when your job runs it is the only job running on that particular node.
+==== sbatch ====
+The sbatch command is used for submitting jobs to the cluster. sbatch accepts a number of options either from the command line, or (more typically) from a batch script. An example of a Slurm batch script is shown below:
+=== Sample script ===
+Make sure you create a directory in which to deposit the ''%%STDIN%%'', ''%%STDOUT%%'', ''%%STDERR%%'' files.
+   mkdir -p $HOME/slurm/out
+<code>
+#!/bin/bash
+#
+#SBATCH --mail-user=cnetid@cs.uchicago.edu
+#SBATCH --mail-type=ALL
+#SBATCH --output=/home/cnetid/slurm/out/%j.%N.stdout
+#SBATCH --error=/home/cnetid/slurm/out/%j.%N.stderr
+#SBATCH --chdir=/home/cnetid/slurm
+#SBATCH --partition=debug
+#SBATCH --job-name=check_hostname_of_node
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --mem-per-cpu=500
+#SBATCH --time=15:00
+hostname
+</code>
+If any of the above options are unclear as to what they do please check the man page for ''%%sbatch%%''
+  man sbatch
+Make sure to replace all instances of the word ''%%cnetid%%'' with your CNETID.
+=== Submitting job script ===
+Using the above example you will want to place your tested code into a file. 'hostname.job' is the file name in this example.
+<code>
+sbatch hostname.job
+</code>
+You can then check the status via squeue or see the output in the output directory '$HOME/slurm/slurm_out'.
+==== srun ====
+Used to submit a job to the cluster that doesn't necessarily need a script.
+<code>
+user@host:~$ srun -n2 hostname
+slurm2
+slurm2
+</code>
+''%%srun%%'' will remain in the foreground until the job has finished.
+<code>
+user@host:~$ srun -n1 sleep 400
+</code>
+==== squeue ====
+This command will show jobs in the queue.
+<code>
+user@host:~$ squeue
+JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
+     debug    sleep     user  R       0:11      1 research2
+</code>
+==== scancel ====
+Cancel one of your own jobs. Please read the ''%%scancel%%'' manual page (''%%man scancel%%'') as there are many ways of canceling your jobs if they are of any complexity.
+<code>
+scancel 29
+</code>
+==== sinfo ====
+View information about Slurm nodes and partitions.
+The following code block shows the what happens when you run the ''%%sinfo%%'' command. You get a list of 'partitions' on which you can run your code. Each partition is comprised of certain types of nodes. In the case below the default (denoted by a *) is 'debug'. The job time limit is short and is meant only to debug your code. The other partitions will usually have a particular purpose in mind. 'hardware', for example, is to be used if you require direct access to the hardware instead of the KVM layer between the hardware and the OS.
+<code>
+user@host:~$ sinfo
+PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
+debug*       up 1-00:00:00      1    mix slurm1
+fast         up 1-00:00:00      6   idle slurm[9-14]
+general      up 21-00:00:0      6   idle slurm[2-6,8]
+pascal       up 3-00:00:00      1   idle gpu2
+quadro       up 3-00:00:00      1   idle gpu1
+titan        up 3-00:00:00      1    mix gpu3
+</code>
+====== Monitoring Jobs ======
+''%%squeue%%'' and ''%%sacct%%'' are two different commands that allow you to monitor job activity in Slurm. ''%%squeue%%'' is the primary and most accurate monitoring tool since it queries the Slurm controller directly. ''%%sacct%%'' gives you similar information for running jobs, and can also report on previously finished jobs, but because it accesses the Slurm database, there are some circumstances when the information is not in sync with squeue.
+Running ''%%squeue%%'' without arguments will list all currently running jobs. It is more common, though to list jobs for a particular user (like yourself) using the ''%%-u%%'' option...
+  squeue -u cnetid
+or for a particular job id.
+  squeue -j 7894
+====== Interactive Jobs ======
+Though batch submission is the best way to take full advantage of the compute power in the job submission cluster, foreground, interactive jobs can also be run.
+An interactive job differs from a batch job in two important aspects:
+  - The partition to be used is the interact partition
+  - Jobs should be initiated with the srun command instead of sbatch.
+This command:
+   srun -p general --pty --cpus-per-task 1 --mem 500 -t 0-06:00 /bin/bash
+will start a command line shell (''%%/bin/bash%%'') on the 'general' queue with 500 MB of RAM for 6 hours; 1 core on 1 node is assumed as these parameters (''%%-n 1 -N 1%%'') were left out. When the interactive session starts, you will notice that you are no longer on a login node, but rather one of the compute nodes dedicated to this queue. The ''%%--pty%%'' option allows the session to act like a standard terminal.
+====== Job Scheduling ======
+We use a [[http://slurm.schedmd.com/priority_multifactor.html|multifactor]] method of job scheduling. Job priority is assigned by a combination of fair-share, partition priority, and length of time a job has been sitting in the queue. The priority of the queue is the highest factor in the job priority calculation. For certain queues this will cause jobs on lower priority queues which overlap with that queue to be requeued. The second most important factor is fair-share score. You can find a description of how Slurm calculates Fair-share [[http://slurm.schedmd.com/priority_multifactor.html#fairshare|here]]. The third most important is how long you have been sitting in the queue. The longer your job sits in the queue the higher its priority grows. If everyone’s priority is equal then FIFO is the scheduling method. If you want to see what your current priority is just do ''%%sprio -j JOBID%%'' which will show you the calculation it does to figure out your job priority. If you do ''%%sshare -u USERNAME%%'' you can see your current fair-share and usage.((https://rc.fas.harvard.edu/resources/running-jobs))
+We also have backfill turned on. This allows for jobs which are smaller to sneak in while a larger higher priority job is waiting for nodes to free up. If your job can run in the amount of time it takes for the other job to get all the nodes it needs, Slurm will schedule you to run during that period. **This means knowing how long your code will run for is very important and must be declared if you wish to leverage this feature. Otherwise the scheduler will just assume you will use the maximum allowed time for the partition when you run.**((https://rc.fas.harvard.edu/resources/running-jobs))
+====== Array Jobs ======
+Instead of submitting multiple jobs to repeat the same process for different data (e.g. getting results for different datasets for a paper) you can use a ''%%job arrays%%''.
+<code>
+#SBATCH start-finish[:step][%maximum concurrent]
+</code>
+Examples:
+<code>
+#SBATCH --array 0-15         0, 1, ..., 15
+#SBATCH --array 1-3          0, 1, 2, 3
+#SBATCH --array 1,3,4,6      1, 3, 4, 6
+#SBATCH --array 1-8:2        1, 3, 5, 7
+#SBATCH --array 1-10:3%2     1, 5, 9, but the only two of these will ever run concurrently.
+</code>
+You can differentiate the various tasks using the variable ''%%SLURM_ARRAY_TASK_ID%%''.
+<code>
+#!/bin/bash
+#
+#SBATCH --mail-user=cnetid@cs.uchicago.edu
+#SBATCH --mail-type=ALL
+#SBATCH --output=/home/cnetid/slurm/out/%j.%N.stdout
+#SBATCH --error=/home/cnetid/slurm/out/%j.%N.stderr
+#SBATCH --chdir=/home/cnetid/slurm
+#SBATCH --partition=debug
+#SBATCH --job-name=check_hostname_of_node
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --mem-per-cpu=500
+#SBATCH --time=15:00
+#SBATCH --array 1-4
+input=("small_dataset" "medium_dataset" "large_dataset" "huge_dataset")
+./process $input[$SLURM_ARRAY_TASK_ID]
+</code>
+Additionally, tasks can be used to add job dependencies and other fancy features. For more information consult the [[https://slurm.schedmd.com/job_array.html|manual]]
+====== Common Issues ======
+^Error ^What does it mean? ^
+| JOB <jobid> CANCELLED AT <time> DUE TO TIME LIMIT | You did not specify enough time for your job to run. The ''%%-t%%'' flag will allow you to set the time limit.|
+| Job <jobid> exceeded <mem> memory limit, being killed | Your job is attempting to use more memory that you have requested for it. Either increase the amount of memory you have requested or reduce the amount of memory usage your application is trying to use.|
+| JOB <jobid> CANCELLED AT <time> DUE TO NODE FAILURE | There can be many reasons for this message, but most often it means that the node your job was set to run on can no longer be contacted by the the Slurm controller.|
+| error: Unable to allocate resources: More processors requested than permitted | It usually has **nothing** to do with privileges you may or may not have. Rather, it usually means that you have allocated more processors than one compute node actually has. |
+====== Using the GPU ======
+===== GRES Multiple GPU's on one system =====
+GRES: Generic Resource. As of 2018-05-04 these only include GPU's.
+Jobs will not be allocated any generic resources unless specifically requested at job submit time using the ''%%--gres%%'' option supported by the ''%%salloc%%'', ''%%sbatch%%'' and ''%%srun%%'' commands. The option requires an argument specifying which generic resources are required and how many resources. The resource specification is of the form ''%%name[:type:count]%%''. The name is the same name as specified by the GresTypes and Gres configuration parameters. type identifies a specific type of that generic resource (e.g. a specific model of GPU). count specifies how many resources are required and has a default value of 1. For example:
+<code>sbatch --gres=gpu:titan:2 ....</code>
+Jobs will be allocated specific generic resources as needed to satisfy the request. If the job is suspended, those resources do not become available for use by other jobs.
+Job steps can be allocated generic resources from those allocated to the job using the ''%%--gres%%'' option with the ''%%srun%%'' command as described above. By default, a job step will be allocated all of the generic resources allocated to the job. If desired, the job step may explicitly specify a different generic resource count than the job. This design choice was based upon a scenario where each job executes many job steps. If job steps were granted access to all generic resources by default, some job steps would need to explicitly specify zero generic resource counts, which we considered more confusing. The job step can be allocated specific generic resources and those resources will not be available to other job steps. A simple example is shown below.
+==== Ok, but I don't want to read the wall of text above ====
+Fine.
+The ''%%--gres%%'' (man srun) is required if you want to make use of a gpu.
+<code>
+  --gpu=gpu:N    # where 'N' is the number of GPUs requested.
+                 # Please try to limit yourself to one GPU per person.
+</code>
+Example when using tensorflow:
+Given the file ''%%f%%'':
+<code>
+#!/usr/bin/env python3
+from tensorflow.python.client import device_lib
+print(device_lib.list_local_devices())
+</code>
+Here we can see that no GPU was allocated to us because we did not specify the ''%%--gres%%'' option
+<code>
+user@bulldozer:~$ srun -p titan --pty /bin/bash
+user@gpu3:~$ ./f 2>&1 | grep physical_device_desc
+user@gpu3:~$
+</code>
+If we request only 1 GPU.
+<code>
+user@bulldozer:~$ srun -p titan --pty --gres=gpu:1 /bin/bash
+user@gpu3:~$ ./f 2>&1 | grep physical_device_desc
+physical_device_desc: "device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:19:00.0, compute capability: 6.1"
+</code>
+If we request 2 GPUs.
+<code>
+user@bulldozer:~$ srun -p titan --pty --gres=gpu:2 /bin/bash
+user@gpu3:~$ ./f 2>&1 | grep physical_device_desc
+physical_device_desc: "device: 0, name: GeForce GTX 1080 Ti, pci bus id: 0000:19:00.0, compute capability: 6.1"
+physical_device_desc: "device: 1, name: GeForce GTX 1080 Ti, pci bus id: 0000:1a:00.0, compute capability: 6.1"
+</code>
+If we request more GPUs then are available.
+<code>
+kauffman3@bulldozer:~$ srun -p titan --pty --gres=gpu:5 /bin/bash
+srun: error: Unable to allocate resources: Requested node configuration is not available
+</code>
+==== Cool, but how do I know where and what resources are available ====
+Turns out the ''%%sinfo%%'' command is super useful.
+<code>
+$ sinfo -O partition,nodelist,gres,features,available
+PARTITION           NODELIST            GRES                FEATURES            AVAIL
+debug*              slurm1              (null)              (null)              up
+fast                slurm[9-14]         (null)              (null)              up
+general             slurm[2-6,8]        (null)              (null)              up
+pascal              gpu2                gpu:gtx1080:1       'pascal,gtx1080'    up
+quadro              gpu1                gpu:p4000:2         'quadro,p4000'      up
+titan               gpu3                gpu:gtx1080ti:4     'pascal,gtx1080ti'  up
+</code>
+FEATURES: Is actually just an arbitrary string in the configuration file that defines a node. However, techstaff hopes it actually provides some useful info.
+GRES: Don't depend on this being accurate, however it will definitely give you a clue as to how many generic resources are in a partition.
+==== Checking how many Generic RESources are being consumed ====
+Simple use the ''%%-O%%'' option for ''%%squeue%%'' and you can see how many generic resources any particular job is consuming.
+<code>
+$ squeue -O username,nodelist,gres
+USER                NODELIST            GRES
+someusername        gpu3                gpu:1
+otherusername       gpu3                gpu:3
+...
+</code>
+===== Environment Variables =====
+==== CUDA_HOME, LD_LIBRARY_PATH ====
+Please make sure you specify $CUDA_HOME and if you want to take advantage of CUDNN libraries you will need to append /usr/local/cuda-x.x/lib64 to the $LD_LIBRARY_PATH environment variable.
+  cuda_version=11.1
+  export CUDA_HOME=/usr/local/cuda-${cuda_version}
+  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_HOME/lib64
+Currently we support the same versions of CUDA that the latest version of CUDNN supports. This is not written in stone and we can accommodate most other versions if required; just let techstaff know what your needs are.
+==== PATH ====
+You may also need to add the following to your ''%%$PATH%%''
+  export PATH=$PATH:/usr/local/cuda/bin
+==== CUDA_VISIBLE_DEVICES ====
+Do not set this variable. It will be set for you by Slurm.
+The variable name is actually misleading; since it does NOT mean the amount of devices, but rather the physical device number assigned by the kernel (e.g. /dev/nvidia2).
+For example: If you requested multiple gpu's from Slurm (--gres=gpu:2), the CUDA_VISIBLE_DEVICES variable should contain two numbers(0-3 in this case) separated by a comma (e.g. 0,1).
+The numbering is relative and specific to you. For example: two users with one job which require two gpus each could be assigned non-sequential gpu numbers. However CUDA_VISIBLE_DEVICES will look like this for both users: 0,1
+===== Example =====
+This sbatch script will get device information from the installed Tesla gpu.
+<code>
+#!/bin/bash
+#
+#SBATCH --mail-user=cnetid@cs.uchicago.edu
+#SBATCH --mail-type=ALL
+#SBATCH --output=/home/cnetid/slurm/slurm_out/%j.%N.stdout
+#SBATCH --error=/home/cnetid/slurm/slurm_out/%j.%N.stderr
+#SBATCH --workdir=/home/cnetid/slurm
+#SBATCH --partition=gpu
+#SBATCH --job-name=get_tesla_info
+export PATH=$PATH:/usr/local/cuda/bin
+export LD_LIBRARY_PATH=$LD_LIBRARY_PATH=/usr/local/cuda/lib
+cat << EOF > /tmp/getinfo.cu
+#include <stdio.h>
+int main() {
+  int nDevices;
+  cudaGetDeviceCount(&nDevices);
+  for (int i = 0; i < nDevices; i++) {
+    cudaDeviceProp prop;
+    cudaGetDeviceProperties(&prop, i);
+    printf("Device Number: %d\n", i);
+    printf("  Device name: %s\n", prop.name);
+    printf("  Memory Clock Rate (KHz): %d\n",
+           prop.memoryClockRate);
+    printf("  Memory Bus Width (bits): %d\n",
+           prop.memoryBusWidth);
+    printf("  Peak Memory Bandwidth (GB/s): %f\n\n",
+.0*prop.memoryClockRate*(prop.memoryBusWidth/8)/1.0e6);
+  }
+}
+EOF
+/usr/local/cuda/bin/nvcc /tmp/getinfo.cu -o /tmp/a.out
+/tmp/a.out
+rm /tmp/a.out
+rm /tmp/getinfo.cu
+</code>
+==== Output ====
+STDOUT will look something like this:
+<code>
+cnetid@linux1:~$ cat $HOME/slurm/slurm_out/12567.gpu1.stdout
+Device Number: 0
+  Device name: Tesla M2090
+  Memory Clock Rate (KHz): 1848000
+  Memory Bus Width (bits): 384
+  Peak Memory Bandwidth (GB/s): 177.408000
+</code>
+STDERR should be blank.
+====== Feedback ======
+If you feel this documentation is lacking in some way please let techstaff know. Email [[techstaff@cs.uchicago.edu]], call (773-702-1031), or stop by our office (Crerar 357).
+====== More ======
+Sometimes other universities have documentation that is better in some areas.
+  - [[ https://hpcc.usc.edu/support/documentation/slurm/ | USC Slurm Docs ]]
+  - [[ https://nesi.github.io/hpc_training/lessons/maui-and-mahuika/slurm | NESI Slurm Docs ]]