User Jobs

This page presents the features of Mufasa that are most relevant to Mufasa's Job Users. Job Users can submit jobs for execution, cancel their own jobs, and see other users' jobs (but not intervene on them).

System resources subjected to limitations

The hardware resources of Mufasa are limited. For this reason, some of them are subjected to limitations, i.e. (these are SLURM's own terms):

cpu

the number of processor cores that a job can use

mem

the amount of RAM that a job can use

gres

the amount of generic resources that a job can use: in Mufasa, the only resources belonging to this set are the GPUs (the virtual GPUs defined by Nvidia MIG, not the physical GPUs)

These are some of the TRES (Trackable RESources) defined by SLURM. From SLURM's documentation: "A TRES is a resource that can be tracked for usage or used to enforce limits against."

SLURM provides jobs with access to resources only for a limited time: i.e., execution time is itself a limited resource.

When a resource is limited, a job cannot use arbitrary quantities of it. On the contrary, the job must specify how much of the resource it requests. Requests are done either by running the job on a partition for which a default amount of resources has been defined, or through the options of the srun command that executes the job via SLURM.

gres syntax

Whenever it is necessary to specify the quantity of gres, i.e. generic resources, a special syntax must be used. In Mufasa gres resources are GPUs, so this syntax applies to GPUs. Number and type of Mufasa's GPUs is described here.

The name of each GPU resource takes the form

Name:Type

where Name is gpu and Type takes the following values:

• 40gb for GPUs with 40 Gbytes of onboard RAM
• 20gb for GPUs with 20 Gbytes of onboard RAM

So, for instance,

gpu:20gb

identifies the resource corresponding to GPUs with 20 GB of RAM. Of this resource Mufasa has a given number, of which a job can request to use some (or all).

When asking for a gres resource (e.g., in an srun command or an SBATCH directive of an execution script), the syntax required by SLURM is

<Name>:<Type>:<quantity>

where quantity is an integer value specifying how many items of the resource are requested. So, for instance, to ask for 2 GPUs of type 20gb the syntax is

gpu:20gb:2

SLURM's generic resources are defined in /etc/slurm/gres.conf. In order to make GPUs available to SLURM's gres management, Mufasa makes use of Nvidia's NVML library. For additional information see SLURM's documentation.

Looking for unused GPUs

GPUs are usually the most limited resource on Mufasa. So, if your job requires a GPU, the best way to get it executed quickly is to request a GPU that is not currently in use.

This command

sinfo -O Gres:100

provides a summary of all the Gres (i.e., GPU) resources possessed by Mufasa. It provides this output:

GRES                                                                                                
gpu:40gb:2(S:0-1),gpu:20gb:3(S:0-1),gpu:10gb:6(S:0-1)

To know which of the GPUs are currently in use, use command

sinfo -O GresUsed:100

which provides an output similar to this:

GRES_USED                                                                                           
gpu:40gb:2(IDX:0-1),gpu:20gb:2(IDX:5,8),gpu:10gb:3(IDX:3-4,6) 

By comparing the two lists (GRES and GRES_USED) above, you can see that at the moment:

• of the 2 40 GB GPUs, both are in use
• of the 3 20 GB GPUs, one is not in use
• of the 6 10 GB GPUs, 3 are not in use

SLURM Partitions

Several execution queues for jobs have been defined on Mufasa. Such queues are called partitions in SLURM terminology. Each partition has features (in term of resources available to the jobs on that queue) that make the partition suitable for a certain category of jobs. SLURM command

sinfo

(link to SLURM docs) provides a list of available partitions. Its output is similar to this:

PARTITION  AVAIL  TIMELIMIT  NODES  STATE NODELIST
debug*        up      20:00      1    mix gn01
small         up   12:00:00      1    mix gn01
normal        up 1-00:00:00      1    mix gn01
longnormal    up 3-00:00:00      1    mix gn01
gpu           up 1-00:00:00      1    mix gn01
gpulong       up 3-00:00:00      1    mix gn01
fat           up 3-00:00:00      1    mix gn01

In this example, available partitions are named “debug”, “small”, “normal”, “longnormal”, “gpu”, “gpulong”, “fat”. The asterisk beside "debug" indicates that this is the default partition, i.e. the one that SLURM selects to run a job when no partition has been specified. (On Mufasa, partition names have been chosen to reflect the type of job that they are dedicated to.)

The columns in the standard output of sinfo shown above correspond to the following information:

PARTITION

name of the partition

AVAIL

state/availability of the partition: see below

TIMELIMIT

maximum runtime of a job allowed by the partition, in format [days-]hours:minutes:seconds

NODES

number of nodes available to jobs run on the partition: for Mufasa, this is always 1 since there is only 1 node in the computing cluster

STATE

state of the node (using these codes); typical values are mixed - meaning that some of the resources of the node are busy executing jobs while other are free, and allocated - meaning that all of the resources of the node are busy

NODELIST

list of nodes available to the partition: for Mufasa this field always contains gn01 since Mufasa is the only node in the computing cluster

One information that the standard output of sinfo doesn't provide is if there are partitions that can only be used by the root user of Mufasa. To know which partiions are root-only, you can use command

sinfo -o "%.10P %.4r"

Its output is

 PARTITION ROOT
    debug*   no
     small   no
    normal   no
longnormal   no
       gpu   no
   gpulong   no
       fat   no

and shows that on Mufasa no partitions are reserved for root.

For what concerns hardware resources (such as CPUs, GPUs and RAM) the amounts of each resource available to Mufasa's partitions are set by SLURM's accounting system, and are not visible to sinfo. See Partition features for a description of these amounts.

Partition features

The output of sinfo (see above) provides a list of available partitions, but (except for time) it does not provide information about the amount of resources that a partition makes available to the user jobs which are run on it. The amount of resources is visible through command

sacctmgr list qos format=name%-10,maxwall,maxtres%-64

which provides an output similar to the following:

Name           MaxWall MaxTRES                                                          
---------- ----------- ---------------------------------------------------------------- 
normal      1-00:00:00 cpu=16,gres/gpu:10gb=0,gres/gpu:20gb=0,gres/gpu:40gb=0,mem=128G  
small         12:00:00 cpu=2,gres/gpu:10gb=1,gres/gpu:20gb=0,gres/gpu:40gb=0,mem=16G    
longnormal  3-00:00:00 cpu=16,gres/gpu:10gb=0,gres/gpu:20gb=0,gres/gpu:40gb=0,mem=128G  
gpu         1-00:00:00 cpu=8,gres/gpu:10gb=2,gres/gpu:20gb=2,mem=64G                    
gpulong     3-00:00:00 cpu=8,gres/gpu:10gb=2,gres/gpu:20gb=2,mem=64G                    
fat         3-00:00:00 cpu=32,gres/gpu:10gb=2,gres/gpu:20gb=2,gres/gpu:40gb=2,mem=256G

Its elements are the following (for more information, see SLURM's documentation):

Name

name of the partition

MaxWall

maximum wall clock duration of the jobs run on the partition (after which they are killed by SLURM), in format [days-]hours:minutes:seconds

MaxTRES

maximum amount of resources ("Trackable RESources") available to a job running on the partition, where

cpu=K means that the maximum number of processor cores is K

gres/gpu:Type=K means that the maximum number of GPUs of class Type (see gres syntax) is K

mem=KG means that the maximum amount of system RAM is K GBytes

Note that there may be additional limits to the possibility to fully exploit the resources of a partition. For instance, there may be a cap on the maximum number of GPUs that can be used at the same time by a single job and/or a single user.

Partition availability

An important information that sinfo provides (column "AVAIL") is the availability (also called state) of partitions. Possible partition states are:

up = the partition is available

Running jobs will be completed

Currently queued jobs will be executed as soon as resources allow

drain = the partition is in the process of becoming unavailable (down)

Running jobs will be completed

Queued jobs will be executed only when the partition becomes available again (up)

down = the partition is unavailable

There are no running jobs

Queued jobs will be executed only when the partition becomes available again (up)

When a partition passes from up to drain no harm is done to running jobs. When a partition passes from any other state to down, running jobs (if they exist) get killed.

A partition in state drain or down requires intervention by a Job Administrator to be restored to up.

Choosing the partition on which to run a job

When launching a job (as explained in Executing jobs on Mufasa) a user should select the partition that is most suitable for it according to the job's features. Launching a job on a partition avoids the need for the user to specify explicitly all of the resources that the job requires, relying instead (for unspecified resources) on the default amounts defined for the partition. Partition features explains how to find out how many of Mufasa's resources are associated to each partition.

The fact that by selecting the right partition for their job a user can pre-define the requirements of the job without having to specify them makes partitions very handy, and avoids possible mistakes. However, users can -if needed- change the resource requested by their jobs wrt the default values associated to the chosen partition. Any element of the default assignment of resources provided by a specific partition can be overridden by specifying an option when launching the job, so users are not forced to accept the default value. However, it makes sense to choose the most suitable partition for a job in the first place, and then to specify the job's requirements only for those resources that have an unsuitable default value.

Resource requests by the user launching a job can be both lower and higher than the default value of the partition for that resource. However, they cannot exceed the maximum value that the partition allows for requests of such resource, if set. If a user tries to run on a partition a job that requests a higher value of a resource than the partition‑specified maximum, the run command is refused.

Tips for partition choice

The larger the fraction of system resources that a job asks for, the heavier the job becomes for Mufasa's limited capabilities. Since SLURM prioritises lighter jobs over heavier ones (in order to maximise the number of completed jobs) it is a very bad idea for a user to ask for their job more resources than it actually needs: this, in fact, will have the effect of delaying (possibly for a long time) job execution. These are tips that you can use to guide partition choice for your job in order to get it executed quickly:

• use the least powerful partition that can support the job
• do not ask for more resources or time than needed
• prefer partitions without access to GPUs
• ask for GPUs that are currently not in use

User limitations on the use of resources

Mufasa is a shared machine, meaning that at any given time its resources subjected to limitations are splitted among all users who request them. This also means that there are limitations on the amount of resources that Mufasa can provide to a given user, whatever the amount of resources that the user requested.

Such limitations come from two sources.

The first source is the fact that each user job is associated to the SLURM partition on which it runs. So, each job can only access the specific subset of resources that are available to the partition.

The second source of limitations is applied by SLURM on a per-user basis. Mufasa is configured in such a way that:

• no more than 2 jobs per user can be running at the same time (note that, since each partition can execute only one job at any given time, the two jobs must make use of different partitions)
• if a user already has a running job, a second job from the same user is only put into execution if there are no requests from other users for the partition it is intended to be run on

Please note that access to some partitions may be restricted to researchers (i.e. M.Sc. students cannot access such partitions).

Running jobs with SLURM: generalities

Note: these are general considerations. See Executing jobs on Mufasa for instructions about running your own processing jobs on Mufasa.

The commands that SLURM provides to run jobs are

srun [options] <command_to_be_run_via_SLURM>

and

sbatch [options] <command_to_be_run_via_SLURM>

(see SLURM documentation: srun, sbatch).

In both cases, <command_to_be_run_via_SLURM> can be any program or Linux shell script. By using srun or sbatch, the command or script specified by <command_to_be_run_via_SLURM> (including any programs launched by it) are added to SLURM's execution queues.

The main difference between srun and sbatch is that the first locks the shell from which it has been launched, so it is only really suitable for processes that use the console for interaction with their user. (You can, though, detach from that shell and come back later using screen.) sbatch, on the other side, does not lock the shell and simply adds the job to the queue, but does not allow the user to interact with the process while it is running.

Additionally, with sbatch <command_to_be_run_via_SLURM> can be an execution script, i.e. a special (and SLURM-specific) type of Linux shell script that includes SBATCH directives. SBATCH directives can be used to specify the values of some of the parameters that would otherwise have to be set using the [options] part of the sbatch command. This is handy because it allows to write down the parameters in an execution script instead of having to write them in the command line while launching a job, which greatly reduces the possibility of mistakes. Also, an execution script is easy to keep and reuse.

The [options] part of srun and sbatch commands is used to tell SLURM the conditions under which it has to execute the job; in particular, it is used to specify what system resources SLURM should reserve for the job.

A quick way to define the set of resources that a program will be provided with is to use SLURM partitions. This is done with option -p <partition_name>. This option specifies that SLURM will run the program on a specific partition, and therefore that it will have access to all and only the resources available to that partition. As a consequence, all options that define how many resources to assign the job will only be able to provide the job with resources that are available to the chosen partition. Jobs that require resources that are not available to the chosen partition do not get executed.

For instance, running

srun -p small ./my_program

makes SLURM run my_program on the partition named “small”. Running the program this way means that the resources associated to this partition will be available to it for use.

Immediately after a srun command is launched by a user, SLURM outputs a message similar to this:

srun: job 10849 queued and waiting for resources

The shell is now locked while SLURM prepares the execution of the user program (if you are using screen you can detach from that shell and come back later).

When SLURM is ready to run the program, it prints a message similar to

srun: job 10849 has been allocated resources

and then executes the program.

Running interactive jobs via SLURM

As explained, SLURM command srun is suitable for launching interactive user jobs, i.e. jobs that use the terminal output and the keyboard to exchange information with a human user. If a user needs this type of interaction, they must run a bash shell (i.e. a terminal session) with a command similar to

srun --pty /bin/bash

and subsequently use the bash shell to run the interactive program. To close the SLURM-spawned bash shell, run (as with any other shell)

exit

Of course, also the “base” shell (i.e. the one that opens when an SSH connection to Mufasa is established) can be used to run programs: however, programs launched this way are not being run via SLURM and are not able to access most of the resources of the machine (in particular, there is no way to make GPUs accessible to them, and they can only access 2 CPUs). On the contrary, running programs with srun or sbatch ensures that they can access all the resources managed by SLURM.

GPU resources (if needed) must always be requested explicitly with parameter --gres=gpu:<20|40>gb:K, where K is an integer between 1 and the maximum number of GPUs of that type available to the partition (see gres syntax). For instance, in order to run an interactive program which needs one GPU we may first run a bash shell via SLURM with command

srun --gres=gpu:20gb:1 --pty /bin/bash

an then run the interactive program from the shell newly opened by SLURM.

A way to specify what resources to assign to the bash shell run via SLURM is to run /bin/bash on one of the available partitions: by doing this, the shell is given access to the default amount of resources associated to the partition. For instance, to run the shell on partition “small” the command is

srun -p small --pty /bin/bash

The general structure of a command requesting SLURM to set up an interactive user job is the following:

srun [‑p <partition_name>] [--job-name=<jobname>] [‑‑gres=<gpu_resources>] [‑‑mem=<mem_resources>] [‑‑cpus‑per‑task=<cpu_amount>] [‑‑time=<duration>] ‑‑pty /bin/bash

Below, the elements of this command are explained.

‑p <partition_name>

specifies the SLURM partition on which the job will be run. If it is not specified, the default partition is used.

Important! The chosen partition limits the resources that can be requested, since it is not allowed to request resources (type or quantity) that exceed what is allowed by the chosen partition.

Important! If ‑‑p <partition_name> is used, options that specify how many resources to assign to the job (such as ‑‑mem=<mem_resources>, ‑‑cpus‑per‑task=<cpu_amount> or ‑‑time=<duration>) can be omitted, greatly simplifying the command. If an explicit amount is not requested for a given resource, the job is assigned the default amount of the resource (as defined by the chosen partition). A notable exception concerns option ‑‑gres=<gpu_resources>, which is always required (see below) if the job needs access to GPUs.

--job-name=<jobname>

Specifies a name for the job. The specified name will appear along with the JOBID number when querying running jobs on the system with squeue. The default job name (i.e., the one assigned to the job when --job-name is not used) is the executable program's name.

‑‑gres=<gpu_resources>

specifies what GPUs to assign to the container. gpu_resources is a comma-delimited list where each element has the form gpu:<Type>:<amount>, where <Type> is one of the types of GPU available on Mufasa (see gres syntax) and <amount> is an integer between 1 and the number of GPUs of such type available to the partition. For instance, <gpu_resources> may be gpu:40gb:1,gpu:10gb:3, corresponding to asking for one "full" GPU and 3 "small" GPUs.

Important! The ‑‑gres parameter is mandatory if the job needs to use the system's GPUs. Differently from other resources (where unspecified requests lead to the assignment of a default amount), GPUs must always be explicitly requested.

‑‑mem=<mem_resources>

specifies the amount of RAM to assign to the container; for instance, <mem_resources> may be 200G

‑‑cpus-per-task=<cpu_amount>

specifies how many CPUs to assign to the container; for instance, <cpu_amount> may be 2

‑‑time=<duration>

specifies the maximum time allowed to the job to run, in the format days-hours:minutes:seconds, where days is optional; for instance, <d-hh:mm:ss> may be 72:00:00

‑‑pty

specifies that the job will be interactive (this is necessary when <command_to_run_within_container> is /bin/bash: see Running interactive jobs via SLURM)

Mufasa is configured to show, as part of the command prompt of a bash shell run via SLURM, a message such as (SLURM ID xx) (where xx is the ID of the /bin/bash process within SLURM). When you see this message, you know that the bash shell you are interacting with is a SLURM-run one.

Another way to know if the current shell is the “base” shell or one run via SLURM is to execute command

echo $SLURM_JOB_ID

If no number gets printed, this means that the shell is the “base” one. If a number is printed, it is the SLURM job ID of the /bin/bash process.

Guidelines you must follow when executing jobs on Mufasa

Mufasa is a shared machine. In order to make best and fair use of its shared resources, every user must carefully follow the three guidelines below:

Limit resource requests to the amount that your job actually needs

An example about CPUs: to make use of multiple CPUs, a process must explicitly support multiple workers/processes in parallel. So if you ask SLURM for 16 CPUs but your code only includes one worker, just one CPU will be used to run it, while the other 15 CPUs will sit idle. Nonetheless, you reserved 16 CPUs: so SLURM will prevent anyone else from using those 15 idle CPUs (until your job terminates).

So it is necessary to adapt every srun or sbatch request to the code that will actually be run. If you don't know how many CPUs your code uses, just run a short-duration job as a test: since SLURM prioritises shorter processes, the test job should get executed much faster than the "real" job.

Run every resource-intensive process via SLURM, i.e. with commands srun or sbatch

If you need to launch resource-intensive programs manually via a shell, simply

1. use SLURM to create an interactive shell
2. when the SLURM-created shell is active, launch your programs from it

Running your heavy processes via SLURM is necessary to ensure that you use Mufasa's resources correctly and respectfully towards the other users: in fact SLURM ensures that Mufasa's resources are distributed to users fairly.

When a SLURM job is not needed anymore, close it with scancel

It is typical that one doesn't know how long a piece of code will take to complete its work. So please make sure to check from time to time if that happened, and -if there's still time before the duration of your SLURM job ends- just scancel the job.

Other resources

The contents of this wiki are specifically tailored for users of Mufasa. They should include everything Mufasa users need to make good use of the machine. However, specific needs vary and advanced users may require advanced functionalities of SLURM that are not covered here.

There are a lot of resources on the internet dealing with the execution of jobs using SLURM. Usually these have been published for the benefit of the users of a specific High Performance Computing system, so there's no guarantee that whatever they suggest will work on Mufasa. If you feel the need to look for external resources, we you may start with this one, which has been prepared by the same people who built Mufasa.

Executing jobs on Mufasa

The main reason for a user to interact with Mufasa is to execute jobs that require resources not available to standard desktop-class machines. Therefore, launching jobs is the most important operation for Mufasa users: what follows explains how it is done.

Important! When launching jobs, always follow the guidelines.

Considering that all computation on Mufasa must occur within Docker containers, the jobs run by Mufasa users are always containers except for menial, non-computationally intensive jobs. This wiki includes directions about preparing Docker containers.

The standard process of launching a user job on Mufasa involves the following steps:

Step 1 --- Use SLURM to run the Docker container where the job will take place

[for interactive and non-interactive user jobs]

Step 2 --- Manually launch the user job from within the container

[for interactive user jobs only]

Interactive and non-interactive user jobs

Interactive user jobs

are jobs that require interaction with the user while they are running, via a bash shell running within the Docker container. The shell is used to receive commands from the user and/or print output messages. For interactive user jobs, the job is usually launched manually by the user (with a command issued via the shell) after the Docker container is in execution.

Non-interactive user jobs

are the most common variety. The user prepares the Docker container in such a way that, when in execution, the container autonomously puts the user's jobs into execution. The user does not have any communication with the Docker container while it is in execution.

Both interactive and non-interactive user jobs can be run via a (quite complex) command directly issued from the terminal opened via SSH. To reduce the possibility of mistakes, it is usually preferable to define an execution script that takes care of launching the job.

Job output

The whole point of running a user job is to collect its output. Usually, such output takes the form of one or more files generated within the filesystem of the Docker container.

As explained below, SLURM includes a mechanism to mount a part of Mufasa's own filesystem onto the container's filesystem: so when the job running within the container writes to this mounted part, it actually writes to Mufasa's filesystem. This means that when the Docker container ends its execution, its output files persist in Mufasa's filesystem (usually in a subdirectory of the user's own /home directory) and can be retrieved by the user at a later time.

The same mechanism can be used to allow user jobs running into a Docker container to read their input data from Mufasa's filesystem (usually a subdirectory of the user's own /home directory).

Using SLURM to run a Docker container

The first step to run a user job on Mufasa is to run the Docker container where the job will take place. A container is a “sandbox” containing the environment where the user's application operates. Parts of Mufasa's filesystem can be made visible (and writable, if they belong to the user's /home directory) to the environment of the container. This allows the containerized user application to read from, and write to, Mufasa's filesystem: for instance, to read data and write results. This wiki includes directions about preparing Docker containers

Each user is in charge of preparing the Docker container(s) where the user's jobs will be executed. In most situations the user can simply select a suitable ready-made container from the many which are already available for use.

In order to run a Docker container via SLURM, a user must use a command similar to the following ones:

For interactive user jobs (parts within [square brackets] are optional):

srun [‑p <partition_name>] ‑‑container-image=<container_path.sqsh> [--job-name=<jobname>] [‑‑no‑container‑entrypoint] ‑‑container‑mounts=<mufasa_dir>:<docker_dir> [‑‑gres=<gpu_resources>] [‑‑mem=<mem_resources>] [‑‑cpus‑per‑task=<cpu_amount>] [‑‑time=<duration>] ‑‑pty /bin/bash

The srun command above runs the Docker Container and opens an interactive shell within the container's environment.

For non-interactive user jobs (parts within [square brackets] are optional):

srun [‑p <partition_name>] ‑‑container-image=<container_path.sqsh> [--job-name=<jobname>] [‑‑no‑container‑entrypoint] ‑‑container‑mounts=<mufasa_dir>:<docker_dir> [‑‑gres=<gpu_resources>] [‑‑mem=<mem_resources>] [‑‑cpus‑per‑task=<cpu_amount>] [‑‑time=<duration>] [<command_to_run_within_container>]

Below, the elements of these commands are explained.

‑p <partition_name>

specifies the SLURM partition on which the job will be run. If it is not specified, the default partition is used.

Important! The chosen partition limits the resources that can be requested, since it is not allowed to request resources (type or quantity) that exceed what is allowed by the chosen partition.

Important! If ‑‑p <partition_name> is used, options that specify how many resources to assign to the job (such as ‑‑mem=<mem_resources>, ‑‑cpus‑per‑task=<cpu_amount> or ‑‑time=<duration>) can be omitted, greatly simplifying the command. If an explicit amount is not requested for a given resource, the job is assigned the default amount of the resource (as defined by the chosen partition). A notable exception concerns option ‑‑gres=<gpu_resources>, which is always required (see below) if the job needs access to GPUs.

--job-name=<jobname>

Specifies a name for the job. The specified name will appear along with the JOBID number when querying running jobs on the system with squeue. The default job name (i.e., the one assigned to the job when --job-name is not used) is the executable program's name.

‑‑container-image=<container_path.sqsh>

specifies the container to be run

‑‑no‑container‑entrypoint

specifies that the entrypoint defined in the container image should not be executed (ENTRYPOINT in the Dockerfile that defines the container). The entrypoint is an element of a Docker container: a command that gets executed as soon as the container is in execution. Option ‑‑no‑container‑entrypoint is useful when -for some reason- the user does not want the entrypoint in the container to be run.

‑‑container‑mounts=<mufasa_dir>:<docker_dir>

specifies what parts of Mufasa's filesystem will be available within the container's filesystem, and where they will be mounted. This is necessary to let the container get input data from Mufasa and/or write output data to Mufasa. For instance, if <mufasa_dir>:<docker_dir> takes the value /home/mrossi:/data this tells srun to mount Mufasa's directory /home/mrossi in position /data within the filesystem of the Docker container. When the docker container reads or writes files in directory /data of its own (internal) filesystem, what actually happens is that files in /home/mrossi get manipulated instead. /home/mrossi is the only part of the filesystem of Mufasa that is visible to, and changeable by, the Docker container.

‑‑gres=<gpu_resources>

specifies what GPUs to assign to the container. gpu_resources is a comma-delimited list where each element has the form gpu:<Type>:<amount>, where <Type> is one of the types of GPU available on Mufasa (see gres syntax) and <amount> is an integer between 1 and the number of GPUs of such type available to the partition. For instance, <gpu_resources> may be gpu:40gb:1,gpu:10gb:3, corresponding to asking for one "full" GPU and 3 "small" GPUs.

Important! The ‑‑gres parameter is mandatory if the job needs to use the system's GPUs. Differently from other resources (where unspecified requests lead to the assignment of a default amount), GPUs must always be explicitly requested.

‑‑mem=<mem_resources>

specifies the amount of RAM to assign to the container; for instance, <mem_resources> may be 200G

‑‑cpus-per-task=<cpu_amount>

specifies how many CPUs to assign to the container; for instance, <cpu_amount> may be 2

‑‑time=<duration>

specifies the maximum time allowed to the job to run, in the format days-hours:minutes:seconds, where days is optional; for instance, <d-hh:mm:ss> may be 72:00:00

‑‑pty

specifies that the job will be interactive (this is necessary when <command_to_run_within_container> is /bin/bash: see Running interactive jobs via SLURM)

<command_to_run_within_container>

the command that will be put into execution within the Docker container as soon as it the container is active. Note that this is mandatory for non-interactive user jobs and optional for interactive user jobs. If specified, this command will be executed in the environment created by Docker.

For interactive user jobs, a typical value for <command_to_run_within_container> is /bin/bash. This instructs srun to open an interactive shell session (i.e. a command-line terminal interface) within the container, from which the user will then run their job. Another typical value for <command_to_run_within_container> is python, which launches an interactive Python session from which the user will then run their job.

For non-interactive user jobs, using [command_to_run_within_container] is one of the two available methods to run the program(s) that the user wants to be executed within the Docker container. The other available method to run the user job(s) is to use the entrypoint of the container. The use of [command_to_run_within_container] is therefore optional.

Using execution scripts to run jobs

The srun commands described in Using SLURM to run a Docker container are very complex, and it's easy to forget some option or make mistakes while using them. For non-interactive jobs, there is a solution to this problem.

When the user job is non-interactive, in fact, the srun command can be substituted with a much simpler sbatch command. As already explained, sbatch can make use of an execution script to specify all the parts of the command to be run via SLURM. So the command to run the Docker container where the user job will take place becomes

sbatch <execution_script>

An execution script is a special type of Linux script that includes SBATCH directives. SBATCH directives are used to specify the values of the parameters that are otherwise set in the [options] part of an srun command.

Note on Linux shell scripts

A shell script is a text file that will be run by the bash shell. In order to be acceptable as a bash script, a text file must: have the “executable” flag set have #!/bin/bash as its very first line Usually, a Linux shell script is given a name ending in .sh, such as my_execution_script.sh, but this is not mandatory. Within any shell script, lines preceded by # are comments (with the notable exception of the initial #!/bin/bash line). Use of blank lines as spacers is allowed.

An execution script is a Linux shell script composed of two parts:

1. a preamble, composed of directives using which the user specifies the values to be given to parameters, each preceded by the keyword SBATCH
2. [optionally] one or more srun commands that launch jobs with SLURM using the parameter values specified in the preamble

The srun commands are optional because jobs can also be launched by the Docker container's own entrypoint.

Below is an execution script template to be copied and pasted into your own execution script text file.

The template includes all the options already described above, plus a few additional useful ones (for instance, those that enable SLURM to send email messages to the user in correspondence to events in the lifecycle of their job). Information about all the possible options can be found in [SLURM's own documentation].

All the SBATCH directives in the script template below are inactive because commented out. To enable a directive, just uncomment it by removing the leading "#". To make them stand out more visibly, in the template the comments corresponding to actual instructions are in bold.

#!/bin/bash

#----------------start of preamble----------------

#SBATCH ‑p <partition_name>

#SBATCH ‑‑container-image=<container_path.sqsh>

#SBATCH --job-name=<name>

#SBATCH ‑‑no‑container‑entrypoint

#SBATCH ‑‑container‑mounts=<mufasa_dir>:<docker_dir>

#SBATCH ‑‑gres=<gpu_resources>

#SBATCH ‑‑mem=<mem_resources>

#SBATCH ‑‑cpus-per-task=<cpu_amount>

#SBATCH ‑‑time=<d-hh:mm:ss>

# The following directives (not described so far) activate SLURM's email notifications:

# the first specifies where they are sent; the following 3 set up notifications start/end/failure of job execution

#SBATCH --mail-user <email_address>

#SBATCH --mail-type BEGIN

#SBATCH --mail-type END

#SBATCH --mail-type FAIL

#----------------end of preamble----------------

# srun <command_to_run_within_container>

# to run the user job, either uncomment (and personalise) the above srun command or use the entrypoint of the Docker container

Nvidia Pyxis

Some of the options described below are specifically dedicated to Docker containers: these are provided by the Nvidia Pyxis package that has been installed on Mufasa as an adjunct to SLURM. Pyxis allows unprivileged users (i.e., those that are not administrators of Mufasa) to execute containers and run commands within them.

More specifically, options ‑‑container-image, ‑‑no‑container‑entrypoint, ‑‑container-mounts are provided to srun by Pyxis.

See the Nvidia Pyxis github page for additional information about the options that it provides to srun.

Launching a user job from within a Docker container

For interactive user jobs, once the Docker container (run as explained here) is up and running, the user is dropped to the interactive environment specified by <command_to_run_within_container>. This interactive environment can be, for instance, a bash shell or an interactive Python console. Once inside the interactive environment, the user can simply run the required program in the usual way (depending on the type of environment).

Please note that the interactive environment of the Docker container does not have any relation with Mufasa's system. The only contact point is the part of Mufasa's filesystem that has been grafted to the container's filesystem via the ‑‑container‑mounts option of srun. In particular, none of the software packages (such as the Nvidia drivers) installed on Mufasa are available in the container, unless they have been installed in it at preparation time (as explained in Docker), or manually after the container is put in execution.

Also note that, once a Docker container launched with srun is in execution, its own bash shell is completely indistinguishable from the bash shell of Mufasa where the srun command that put the container in execution was issued. The two shells share the same terminal window. The only clue to the fact that you now are, in fact, in the container's shell may be the command prompt, which should now show your location as /opt.

Detaching from a running job with screen

A consequence of the way srun operates is that if you launch an interactive user job, the shell where the command is running must remain open: if it closes, the job terminates. That shell runs in the terminal of your own PC where the SSH connection to Mufasa exists.

If you do not plan to keep the SSH connection to Mufasa open (for instance because you have to turn off or suspend your PC), there is a way to keep your interactive job alive. Namely, you should use command srun inside a screen session (often simply called "a screen"), then detach from the screen (here is one of many tutorials about screen available online).

Once you have detached from the screen session, you can close the SSH connection to Mufasa without damage. When you need to reach your (still running) job again, you can can open a new SSH connection to Mufasa and then reattach to the screen.

A use case for screen is writing your program in such a way that it prints progress advancement messages as it goes on with its work. Then, you can check its advancement by periodically reconnecting to the screen where the program is running and reading the messages it printed.

Basic usage of screen is explained below.

Creating a screen session, running a job in it, detaching from it

1. Connect to Mufasa with SSH
2. From the Mufasa shell, run

   screen

3. In the screen session ("screen") thus created (it has the look of an empty shell), launch your job with srun
4. Detach from the screen by pressing ctrl + A followed by D: you will come back to the original Mufasa shell, while your process will go on running in the screen
5. You can now close the SSH connection to Mufasa without damaging your running job

Reattaching to an active screen session

1. Connect to Mufasa with SSH
2. In the Mufasa shell, run

   screen -r

3. You are now back to the screen where you launched your job

Closing (i.e. destroying) a screen session

When you do not need a screen session anymore:

1. reattach to the screen as explained above
2. destroy the screen by pressing ctrl + A followed by \ (i.e., backslash)

Of course, any program running within the screen gets terminated when the screen is destroyed.

Using salloc to reserve resources

What is salloc?

salloc is a SLURM command that allows a user to reserve a set of resources (e.g., a 40 GB GPU) for a given time in the future.

The typical use of salloc is to "book" an interactive session where the user enjoys complete control of a set of resources. The resources that are part of this set are chosen by the user. Within the "booked" session, any job run by the user that relies on the reserved resources is immediately put into execution by SLURM.

More precisely:

• the user, using salloc, specifies what resources they need and the time when they will need them;
• when the delivery comes, SLURM creates an interactive shell session for the user;
• within such session, the user can use srun and sbatch to run programs, enjoying full (i.e. not shared with anyone else) and instantaneous access to the resources.

Resource reservation using salloc is only possible if the request is done in advance wrt the delivery time. The more the resources that the user wants to reserve are in high demand, the more anticipated the request should be to ensure that SLURM is able to fulfill it.

When a user makes a request for resources with salloc, the request (called an allocation) gets added to the job queue of SLURM of the requisite partition as a job in pending (PD) state (job states are described here). Indeed, resource allocation is the first part of SLURM's process of executing a user job, while the second part is running the program and letting it use the allocated resources. Using salloc actually corresponds to having SLURM perform the first part of the process (resource allocation) while leaving the second part (running programs) to the user.

Until the delivery time specified by the user comes, the allocation remains in state PD, and other jobs requesting the same resources, even if submitted later, are executed. While the request waits for the delivery time, however, it accumulates a priority that increases over time. The longer the allocation stays in the PD state, the stronger this accumulation of priority: so, by requesting resources with salloc well in advance of the delivery time, users can ensure that the resources they need will be ready for them at the requested delivery time, even if these resources are highly contended.

salloc commands

salloc commands use a similar syntax to srun commands. In particular, salloc lets a user specify what resources they need and -importantly- a delivery time for the requested resources (delivery time can also be specified with srun, but in that case it is not very useful).

The typical salloc command has this form:'

salloc [-p <partition_name>] [--job-name=<jobname>] [‑‑gres=<gpu_resources>] [‑‑mem=<mem_resources>] [‑‑cpus‑per‑task=<cpu_amount>] ‑‑time=<duration> --begin=<time>

The parts of the above commands within [square brackets] are optional.

Below, the elements of the command are explained.

‑p <partition_name>

specifies the SLURM partition on which the job will be run. If it is not specified, the default partition is used.

Important! The chosen partition limits the resources that can be requested, since it is not allowed to request resources (type or quantity) that exceed what is allowed by the chosen partition.

Important! If ‑‑p <partition_name> is used, options that specify how many resources to assign to the job (such as ‑‑mem=<mem_resources>, ‑‑cpus‑per‑task=<cpu_amount> or ‑‑time=<duration>) can be omitted, greatly simplifying the command. If an explicit amount is not requested for a given resource, the job is assigned the default amount of the resource (as defined by the chosen partition). A notable exception concerns option ‑‑gres=<gpu_resources>, which is always required (see below) if the job needs access to GPUs.

--job-name=<jobname>

Specifies a name for the job corresponding to the resource allocation. The specified name will appear along with the JOBID number when querying running jobs on the system with squeue. The default job name (i.e., the one assigned to the job when --job-name is not used) is "interact".

‑‑gres=<gpu_resources>

specifies what GPUs are requested. gpu_resources is a comma-delimited list where each element has the form gpu:<Type>:<amount>, where <Type> is one of the types of GPU available on Mufasa (see gres syntax) and <amount> is an integer between 1 and the number of GPUs of such type available to the partition. For instance, <gpu_resources> may be gpu:40gb:1,gpu:10gb:3.

Important! The ‑‑gres parameter is mandatory if the job needs to use the system's GPUs. Differently from other resources (where unspecified requests lead to the assignment of a default amount), GPUs must always be explicitly requested.

‑‑mem=<mem_resources>

specifies the amount of RAM requested; for instance, <mem_resources> may be 200G

‑‑cpus-per-task=<cpu_amount>

specifies how many CPUs are requested; for instance, <cpu_amount> may be 2

‑‑time=<duration>

specifies the maximum time allowed to the job to run, in the format days-hours:minutes:seconds, where days is optional; for instance, <d-hh:mm:ss> may be 72:00:00. While the interactive session associated to the allocation is active, the user can decide to cancel the allocation at any time just by closing the session (e.g., with command exit for bash)

--begin=<time>

specifies the delivery time of the resources reserved with salloc, according to the syntax described below. The delivery time must be a future time.

Syntax of parameter --begin

If the allocation is for the current day, you can specify <time> as hours and minutes in the form

HH:MM

If you want to specify a time of a different day, the form for is YYYY-MM-DDTHH:MM, where the uppercase 'T' separates date from time.

It is also possible to specify as relative to the current time, in one of the following forms:

now+Kminutes

now+Khours

now+Kdays

where K is a (positive) integer.

Examples:

--begin=16:00

--begin=now+1hours

--begin=now+1days

--begin=2030-01-20T12:34:00

Note that Mufasa's time zone is GMT, so <time> must be expressed in GMT as well. If you want to know Mufasa's current time, use command

date

It provides an output similar to the following:

Thu Nov 10 16:43:30 UTC 2022

How to use salloc

In the typical scenario, the user of salloc will make use of screen. Command screen creates a shell session (called "a screen") that it is possible to abandon without closing it ("detaching from the screen"). It is then possible to reach again the screen at a later time ("reattaching to the screen"). This means that a user can create a screen, run salloc within it to create an allocation for time X, detach from the screen and reattach to it just before time X to use the reserved resources from the interactive session created by salloc.

More precisely, the operations needed to do this are the following:

1. Connect to Mufasa with SSH.
2. From the Mufasa shell, run

   screen

3. In the screen session ("screen") thus created run the salloc command, specifying via its options the resources you need and the time at which you want them delivered.
4. SLURM will respond with a message similar to

   salloc: Pending job allocation XXXX

5. Detach from the screen by pressing ctrl + A followed by D: you will come back to the original Mufasa shell.
6. You can now close the SSH connection to Mufasa without damaging your resource allocation request.
7. At the delivery time you specified in the salloc command, connect to Mufasa with SSH.
8. Once you are in the Mufasa shell, reattach to the screen with command

   screen -r

9. You are now back to the screen where you used salloc; as soon as SLURM provides to you with the resources you reserved, message "salloc: Pending job allocation XXXX" changes to the shell prompt.
10. You are now in the interactive shell session you booked with salloc. From here, you can run any programs you want, including srun and sbatch. For the whole duration of the allocation, your programs have unrestricted use of all the resources you reserved with salloc.
    Important! Any job run within the shell session is subject to the time limit (i.e., maximum duration) imposed by the partition it is running on! Therefore, if the job reaches the time limit, it gets forcibly terminated by SLURM. Termination depends exclusively from the time limit: so it occurs even if the end time for the allocation has not been reached yet. (Of course, the job also gets terminated if the allocation ends.)
11. Once the interactive shell session is not needed anymore, cancel it by exiting from the session with

    exit

    (Note that if you get to the end of the time period you specified in your request without closing the shell session, SLURM does it for you, killing any programs still running.)
12. You are now back to your screen. Destroy it by pressing ctrl + A followed by \ (i.e., backslash) to get back to the Mufasa shell.

Cancelling a resource request made with salloc

To cancel a request for resources made as explained in How to use salloc, follow these steps:

1. Connect to Mufasa with SSH.
2. Once you are in the Mufasa shell, reattach to the screen where you used command salloc with command

   screen -r

3. You should see the message "salloc: Pending job allocation XXXX" (if the allocation is still pending) or ""salloc: job XXXX queued and waiting for resources" (if the allocation is done and waiting for its start time). Now just press Ctrl + C. This communicates to SLURM your intention to cancel your request for resources.
4. SLURM will communicate the cancellation with message

   salloc: Job allocation XXXX has been revoked.

5. Destroy the screen by pressing ctrl + A followed by \ (i.e., backslash) to get back to the Mufasa shell.

Automatic job caching

When a job is run via SLURM (with or without an execution script), Mufasa exploits a (fully tranparent) caching mechanism to speed up its execution. The speedup is obtained by removing the need for the running job to execute accesses to the (mechanical and therefore relatively slow) HDDs where /home partitions reside, substituting them with accesses to (solid-state and therefore much faster) SSDs.

Each time a job is run via SLURM, this is what happens automatically:

1. Mufasa temporarily copies code and associated data from the directory where the executables are located (in the user's own /home) to a cache space located on system SSDs
2. Mufasa launches the cached copy of the user executables, using the cached copies of the data as its input files
3. The executables create their output files in the cache space
4. When the user jobs end, Mufasa copies the output files from the cache space back to the user's own /home

The whole process is completely transparent to the user. The user simply prepares the executable (or the execution script) in a subdirectory of their /home directory and runs the job. When job execution is complete, the user finds their output data in the origin subdirectory of /home, exactly as if the execution actually occurred there.

Important! The caching mechanism requires that during job execution the user does not modify the contents of the /home subdirectory where executable and data were at execution time. Any such change, in fact, will be overwritten by Mufasa at the end of the execution, when files are copied back from the caching space.

Monitoring and managing jobs

SLURM provides Job Users with tools to inspect and manage jobs. While a Job User is able to see all users' jobs, they are only allowed to interact with their own.

The main commands used to interact with jobs are squeue to inspect the scheduling queues and scancel to terminate queued or running jobs.

Inspecting jobs with squeue

Running command

squeue

provides an output similar to the following:

JOBID PARTITION     NAME     USER ST       TIME  NODES NODELIST(REASON)
  520       fat     bash acasella  R 2-04:10:25      1 gn01
  523       fat     bash amarzull  R    1:30:35      1 gn01
  522       gpu     bash    clena  R   20:51:16      1 gn01

This output comprises the following information:

JOBID

Numerical identifier of the job assigned by SLURM

This identifier is used to intervene on the job, for instance with scancel

PARTITION

the partition that the job is run on

NAME

the name assigned to the job; can be personalised using the --job-name option

USER

username of the user who launched the job

ST

job state (see Job state for further information)

TIME

time that has passed since the beginning of job execution

NODES

number of nodes where the job is being executed (for Mufasa, this is always 1 as it is a single machine)

NODELIST (REASON)

name of the nodes where the job is being executed: for Mufasa it is always gn01, which is the name of the node corresponding to Mufasa.

To limit the output of squeue to the jobs owned by user <username>, it can be used like this:

squeue -u <username>

Interpreting Job state as provided by squeue

Jobs typically pass through several states in the course of their execution. Job state is shown in column "ST" of the output of squeue as an abbreviated code (e.g., "R" for RUNNING).

The most relevant codes and states are the following:

PD PENDING

Job is awaiting resource allocation.

R RUNNING

Job currently has an allocation.

S SUSPENDED

Job has an allocation, but execution has been suspended and CPUs have been released for other jobs.

CG COMPLETING

Job is in the process of completing. Some processes on some nodes may still be active.

CD COMPLETED

Job has terminated all processes on all nodes with an exit code of zero.

Beyond these, there are other (less frequent) job states. The SLURM doc page for squeue provides a complete list of them.

Knowing when jobs are expected to end or start

If you are interested in understanding when jobs are expected to start or end, use command

squeue -o "%5i %8u %10P %.2t |%19S |%.11L|"

which provides an output is similar to the following:

JOBID USER     PARTITION  ST |START_TIME          |  TIME_LEFT|
5307  thuynh   fat        PD |2022-11-11T17:55:54 | 3-00:00:00|
5308  thuynh   fat        PD |2022-11-11T17:55:54 | 3-00:00:00|
5296  cziyang  fat         R |2022-11-08T16:58:03 | 1-00:48:14|
5306  thuynh   fat         R |2022-11-10T08:13:30 | 2-16:03:41|
5297  gnannini fat         R |2022-11-08T17:55:54 | 1-01:46:05|
5336  ssaitta  gpu         R |2022-11-10T08:13:00 |    6:03:11|
5358  dmilesi  gpulong     R |2022-11-10T15:11:32 | 2-23:01:43|
5338  cziyang  gpulong     R |2022-11-10T09:45:01 | 1-17:35:12|

For running jobs (state R)

column "START_TIME" tells you when the job started its execution

column "TIME_LEFT" tells you how much remains of the running time requested by the job

For pending jobs (state PD)

column "START_TIME" tells you when the job is expected to start its execution

column "TIME_LEFT" tells you how much running time has been requested by the job

Important! Start and end times are forecasts based on the features of current jobs in the queues, and may change if running jobs end prematurely and/or if new jobs with higher priority are added to the queues. So these times should never be considered as certain.

If you simply want to know when pending jobs (state PD) are expected to begin execution, use

squeue --start

which lists pending jobs in order of increasing START_TIME (the job on top is the one which will be run first). For each pending job the command provides an output similar to the example below:

JOBID PARTITION     NAME     USER ST          START_TIME  NODES SCHEDNODES           NODELIST(REASON)
 5090       fat training   thuynh PD 2022-10-27T09:28:01      1 (null)               (Resources)

Getting detailed information about a job

If needed, complete information about a job (either pending or running) can be obtained using command

scontrol show job <JOBID>

where <JOBID> is the number from the first column of the output of squeue. The output of this command is similar to the following:

JobId=936 JobName=bash
   UserId=acasella(1001) GroupId=acasella(1001) MCS_label=N/A
   Priority=7885 Nice=0 Account=research QOS=normal
   JobState=RUNNING Reason=None Dependency=(null)
   Requeue=0 Restarts=0 BatchFlag=0 Reboot=0 ExitCode=0:0
   RunTime=03:21:59 TimeLimit=3-00:00:00 TimeMin=N/A
   SubmitTime=2022-02-08T11:57:24 EligibleTime=2022-02-08T11:57:24
   AccrueTime=Unknown
   StartTime=2022-02-08T11:57:24 EndTime=2022-02-11T11:57:24 Deadline=N/A
   PreemptEligibleTime=2022-02-08T11:57:24 PreemptTime=None
   SuspendTime=None SecsPreSuspend=0 LastSchedEval=2022-02-08T11:57:24 Scheduler=Main
   Partition=fat AllocNode:Sid=rk018445:4034
   ReqNodeList=(null) ExcNodeList=(null)
   NodeList=gn01
   BatchHost=gn01
   NumNodes=1 NumCPUs=8 NumTasks=1 CPUs/Task=8 ReqB:S:C:T=0:0:*:*
   TRES=cpu=8,mem=128G,node=1,billing=8,gres/gpu:40gb=1
   Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*
   MinCPUsNode=8 MinMemoryNode=128G MinTmpDiskNode=0
   Features=(null) DelayBoot=00:00:00
   OverSubscribe=YES Contiguous=0 Licenses=(null) Network=(null)
   Command=/bin/bash
   WorkDir=/home/acasella
   Power=
   TresPerNode=gres:gpu:40gb:1

In particular, the line beginning with "StartTime=" provides expected times for the start and end of job execution. As explained in Knowing when jobs are expected to end or start, start time is only a prediction and subject to change.

Canceling a job with scancel

It is possible to cancel a job using command scancel, either while it is waiting for execution or when it is in execution (in this case you can choose what system signal to send the process in order to terminate it). The following are some examples of use of scancel adapted from SLURM's documentation.

scancel <JOBID>

removes queued job <JOBID> from the execution queue.

scancel --signal=TERM <JOBID>

terminates execution of job <JOBID> with signal SIGTERM (request to stop).

scancel --signal=KILL <JOBID>

terminates execution of job <JOBID> with signal SIGKILL (force stop).

scancel --state=PENDING --user=<username> --partition=<partition_name>

cancels all pending jobs belonging to user <username> in partition <partition_name>.

Knowing what jobs you ran today

Command

sacct -X

provides a list of all jobs run today by your user.