Child pages
  • Mox_scheduler
Skip to end of metadata
Go to start of metadata

You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 73 Next »

 

This article is for mox.hyak (hyak nextgen) and for ikt.hyak (hyak classic).

Mox and ikt use a scheduler called slurm. It is similar to but different from the PBS based scheduler used earlier on hyak classic. 

Below xyz is your hyak group name and abc is your UW netid.

To logon mox:

ssh abc@mox.hyak.uw.edu

To logon ikt:

ssh abc@hyak.washington.edu

The above command gives you access to the login node of mox. The login node is only for logging in and submitting jobs. The computational work is  done on a compute node. As shown below, you can get either an interactive compute node or submit a batch job. The build node is a special compute node which can connect to the internet.

To see information about your partition (aka allocations):

sinfo -p xyz

To see the usage of your group's nodes:

squeue -p xyz

To see your jobs:

squeue -u abc

Below mox specific command  shows  the number of nodes etc. of all allocations.

hyakalloc

Interactive Single Node Usage with srun:

The build node can connect to outside mox. It is useful for using git, transferring files to outside mox or getting files from outside mox, installing packages in R or Python etc.

To get an interactive build node for 2 hours:

For mox.hyak:

srun -p build --time=2:00:00 --mem=100G --pty /bin/bash

For ikt.hyak:

srun -p build --time=2:00:00 --mem=40G --pty /bin/bash

(Note: --pty /bin/bash must be the last option in above command.)

 

Specifying memory:

           (1) It is important to specify --mem=100G. If this is not specified then the the SLURM scheduler limits the usage of memory to

                a default value even if more memory is available on the node. Usually the default value is quite low. )

           (2) For ikt.hyak regular compute node use --mem=60G . For ikt.hyak for build node use --mem=40G.

           (3) If you know that your group has larger memory nodes, then you can use larger values in --mem option.

                The value in --mem option should be smaller than the memory on the node. This is because the operating system also needs some memory.

              For 64GB nodes, use --mem=60G,   

              For 128GB nodes, use ---mem=120G

              For 256GB nodes, use --mem=248G

              For 512GB nodes, use --mem=500G

 

An interactive node in your own group cannot connect to outside mox.

To get an interactive node in your own group for 2 hours:

srun -p xyz -A xyz --time=2:00:00 --mem=100G --pty /bin/bash

On mox and ikt, the -p and -A options should be the same. (Except when you are using the build node with "-p build", then you do not need to give the -A option.)

Issue below comand at an interactive node prompt to find the list of SLURM environment variables:

export | grep SLURM


Interactive Multiple Node Usage with srun:

If you are setting up an application which uses multiple nodes (e.g. Apache Spark), you will need interactive access to multiple nodes .

To get 2 nodes for interactive use:

srun -N 2 -p xyz -A xyz  --time=2:00:00 --mem=100G --pty /bin/bash

When the above command runs, then you will have been allocated 2 nodes but will be on one of the two allocated node. The --mem option is for memory per node.

In order to find the names of the nodes that you have been allocated, issue below command:

scontrol show hostnames

Once you know the node names, then you can use them for your work (e.g. Apache Spark, Hadoop etc.)

Below command will tell you about other SLURM environment variables.

export | grep SLURM

 

scancel

Use scancel to cancel jobs. For example to cancel a job with job id 1234:

scancel 1234

If your userid is abc then you can cancel all your jobs by using below command:

scancel -u abc

sstat

Use sstat for information (e.g. memory usage) about your currently running jobs.

sstat --format=jobid,MaxVMSize

For details of other format options see https://slurm.schedmd.com/sstat.html

sacct

Use sacct for information (e.g. memory usage) about your jobs which have completed.

sacct --format=jobid,elapsed,MaxVMSize

For details of other format options see https://slurm.schedmd.com/sacct.html

sreport

For historical usage for group xyz for month of March 2018:

sreport cluster UserUtilizationByAccount Start=2018-03-01 End=2018-03-31 | grep xyz

 

Batch usage:

Submit a batch job from mox login node:

sbatch -p xyz -A xyz myscript.slurm

The script myscript.slurm is similar to myscript.pbs used in hyak classic. Below is an example slurm script. The --mem option is for memory per node.

#!/bin/bash

## Job Name

#SBATCH --job-name=myjob

## Allocation Definition

## On mox and ikt, the account and partition options should be the same.
#SBATCH --account=xyz
#SBATCH --partition=xyz

## Resources

## Nodes

#SBATCH --nodes=1   

## Walltime (3 hours). Do not specify a walltime substantially more than your job needs.

#SBATCH --time=3:00:00

## Memory per node. It is important to specify the memory since the default memory is very small.

## For mox, --mem may be more than 100G depending on the memory of your nodes.

## For ikt, --mem may be 60G or more depending on the memory of your nodes.

## See above section on "Specifying memory" for choices for --mem.

#SBATCH --mem=100G

## Specify the working directory for this job

#SBATCH --workdir=/gscratch/xyz/abc/myjobdir

##turn on e-mail notification

#SBATCH --mail-type=ALL

#SBATCH --mail-user=your_email_address

myprogram

 

 

More details are here:

Hyak mox Overview

https://slurm.schedmd.com/quickstart.html

https://slurm.schedmd.com/documentation.html

 

 

===== Below is for Advanced users only ====

Below is about salloc. Do not use salloc unless you have a specific reason.


Interactive Multiple Node Usage with salloc:

If you are setting up an application which uses multiple nodes (e.g. Apache Spark), you will need interactive access to multiple nodes .

To get 2 nodes for interactive use:

salloc -N 2 -p xyz -A xyz  --time=2:00:00 --mem=100G

When the above command runs, then you will have been allocated 2 nodes but will still be on the mox login node.

If you issue a command like below then srun will run the command hostname on each node:

srun hostname

In order to find the names of the nodes that you have been allocated, issue below command:

scontrol show hostnames

Once you know the node names, then you can use them for your work (e.g. Apache Spark, Hadoop etc.)

Below command will tell you about other SLURM environment variables.

export | grep SLURM

 

srun vs salloc

If no nodes have been allocated then (1) and (2) are equivalent.

(1) srun -p xyz -A xyz --time=4:00:00  --mem=100G --pty /bin/bash

(2) salloc -p xyz -A xyz --time=4:00:00   --mem=100G          This allocates the node.

     srun --pty /bin/bash            This uses the same node.

If nodes have already been allocated by using salloc then srun just uses those nodes.

Alternative to srun is to allocate a node and then ssh to it e.g.

Allocate a node:

salloc -p xyz -A xyz --time=4:00:00 --mem=100G

Find the name of the allocated node:

export | grep LIST

If above command gives a hostname of n1234 then you can ssh to n1234 and use it for your work.

If you you use srun to get an interactive node then when you exit the node, your access to the node also ends.

If you use salloc followed by ssh to get an interactive node then when you exit the node, the node will still be allocated to you. Hence, you can again ssh to the node till the till the time in the salloc command is over.

 

 

  • No labels