User Tools

Site Tools


reserve

Book the testbed with OAR

Once you have accessed the testbed (i.e. connected on CorteXlab server), you can reserve the nodes for an experimentation.

To avoid cross interference between multiple experiments, only one person can use the whole CorteXlab testbed at a time. The OAR scheduler is used to book nodes on the platform. As soon as you book one or more nodes, the CorteXlab room is reserved for your usage during the requested time.

The state of reservation of the CorteXlab testbed can be visualized here: http://xp.cortexlab.fr/drawgantt/

The role of OAR is to schedule node reservations. It manages jobs associated with users. A job has a start time, a duration (walltime), and uses some resources (CorteXlab nodes).

Submissions

The principle of operation of CorteXlab is that users submit jobs to OAR. When the job starts, the user gets exclusive access to the platform, and inside an OAR job, the user can perform (interactively, or in batch) one or several experiments.

A basic example to submit an OAR interactive job requesting all available nodes, for the default duration (which is 2 hours):

$ oarsub -I -l nodes=BEST

The same example with a max duration of 4 hours :

$ oarsub -I -l nodes=BEST,walltime=4:00:00

This command will wait for the resources to be available, and as soon as they are (i.e. -I stands for interactive), a job is allocated, is started, and a subshell is instantiated where you can work on experiments. As soon as the subshell is closed, the job ends. (It can be useful to work in a screen session to avoid losing jobs in case of network disconnection).

Submissions may not be interactive. You can provide a script name to execute when the job starts. It has the strong advantage that it allows you to avoid waiting for the job start, which can be long if the platform is heavily used. But for this to work you have to automate everything:

$ oarsub -l nodes=BEST '/path/to/script/to/execute/when/job/starts script args'

A particular case of this syntax is:

$ oarsub -l nodes=BEST 'sleep 1000000'

It allows you to have a job which is not tied to a terminal, but you still need to manually submit minus tasks when the job starts

Reservations

By default OAR submissions are scheduled as soon as possible. It is also possible to ask for an OAR reservation where you choose the date at which the job will be scheduled.

This other simple example is reserving all the nodes on the 18 of September 2015 from 10AM to 11AM:

$ oarsub -l nodes=BEST,walltime=1:00:00 -r "2015-09-18 10:00:00" 

Booking specific nodes

If you want to reserve specific nodes, there are several possible syntaxes.

To make a submission using two nodes:

$ oarsub -I -l nodes=2

But the nodes will be randomly chosen by OAR, so you'll have to adapt your task's scenario to the allocated nodes.

It is possible to ask for explicit nodes with a less user-friendly syntax (especially in situations where you need lots of nodes). For example, to make a submission using specifically nodes 4 and 6, for a 30 minutes job:

$ oarsub -l {"network_address in ('mnode4.cortexlab.fr', 'mnode6.cortexlab.fr')"}/nodes=2,walltime=0:30:00

Despite this syntax being not user-friendly, we strongly encourage you to use it, since it has many advantages:

  • Nodes are often shutdown, in energy saving mode. Booking only the needed nodes ensures that only these ones will be wakeup. This contributes to increase node lifetime and saving energy
  • As only needed nodes are asked, it avoids your oar job being canceled or postponed in case one node that you don't use is unavailable
  • It allows tracing more accurately resources usage

Booking the room without any node

You can book the room only, without any node with the following syntax:

$ oarsub -t noop -l {"type='cortexlab-room'"}/nodes=1 "sleep infinity"

This is useful in particular if you want to go physically in the room for experimenting with specific hardware, because it avoids waking any node if they are in energy saving mode.

A note on OAR job scheduling

Be aware that OAR behaviour may sometimes be counter-intuitive: You may think that just because you ran oarsub successfuly and it returns a job id, your job is running, but this assumption is wrong: OAR tells you that it accepts your job submission but it may schedule it later for various reasons:

  • because the nodes are currently shutdown, so it needs to wake them up, which may take some time
  • because another job is running
  • because the resources you ask are currently not available but OAR expects them to be available in the future
  • etc.

So, the only reliable way to be sure that your job is actually running is to check that the job's state is “Running” with command:

oarstat -fj <JOBID>

Tasks submitted to minus will never start unless the job is “Running” anyway.

To sum-up things:

  • Submissions versus Reservations:
    • Submission: To get the resources as soon as possible. You do not control when the job will be scheduled, and as long as the job hasn't started, the schedule may change. You may get the resources right now if OAR can (and decides to) schedule the job immediately, but there's no guarantee. With a Submission you are sure to get the resources you asked, but you don't control when (the extreme case is that if a resource becomes permanently unavailable, like for example when a node is broken, then your job will never be scheduled, ie. it will stay in “Waiting” state indefinitely)
    • Reservation; you ask for resources at a specific date. With a Reservation you are sure to get the resources at the date you requested (with sometimes a few minutes margin), but you are not sure to get exactly the resources you requested (some resources may have become unavailable at that date)
  • Interactive versus Non Interactive:
    • in the non interactive case, you provide an executable which will be executed by OAR on airlock during your job. This executable can be anything, a sleep infinity command, a script, a binary executable. It will be run at job start and killed at job end. If it ends before the walltime of the job, the job terminates. You can for example provide a script that runs one or several minus task submit commands to fully automate an experimental campaign (but beware that your script needs to wait for the end of the last minus task execution, otherwise, the job will be killed (together with the last minus task) when the script terminates.
    • in the interactive case, the executable is actually an interactive subshell started by OAR on airlock. Beware, if this subshell terminates, the job will be killed as well (you can prevent that by running interactive submissions inside a gnu screen or tmux to protect from ssh disconnections. Both are installed on airlock. For this, you need to open the ssh connection to airlock, then start a gnu screen / tmux session, then run your interactive jobs in that session. The gnu screen / tmux session will survice ssh disconnections or shutdowns of your workstation, and can be rejoined later)

Note that almost all OAR jobs terminate with status “Error”. It's actually a side effect of the fact that the status of a terminated OAR job depends on the return code of the job executable. If it returns 0 (the unix convention for success, or True) the job is in “Terminated” state, if it returns anything but 0 (the unix convention for failure, or False), the job is in “Error” state. Most jobs use sleep infinity as executable, and sleep returns something different than 0 when it is killed at job end.

A note on energy saving

When nodes are unused, and after a timeout, they will be automatically shutdown (and will appear as “Standby” in the drawgantt)

When a job is submitted, shutdown nodes are waken up. Thus the job will not start immediately, it will wait for the nodes to have started. If some nodes are not started after the timeout, they will be set in state “Absent”. If the job is a reservation, it will start without these nodes. If it is a submission, it will scheduled later (possibly never, OAR will wait for the resource to be back, which may never happen)

When a job is a reservation, OAR knows the scheduled start and starts to wakeup nodes before the scheduled start of the job, so usually the job should start approximately as planned.

Since energy saving is active, it is strongly encouraged to submit/reserve only the nodes you need, and if you don't need any node, to only submit/reserve the room. It will save energy, avoid heating the room needlessly, increase longevity of our hardware. Also nodes are awaken by groups of 5, so the more nodes you reserve, the longer the startup time of your job, maybe up to 15 minutes if you reserve all nodes.

Advanced usage: sharing the platform

It is possible to share the platform, for specific situations such as tutorials, courses, challenges. In these situations, you want several users to be able to use the platform at the same time. For this, you need to follow these steps:

The organizer of the tutorial/course/challenge submits or reserve the whole (or part of) the platform for the duration of the event, with the -t container option:

$ oarsub -l nodes=BEST,walltime=4:00:00 -t container -r '2018-07-21 14:00:00'

This will reserve all available nodes for a 4 hours event, between 14 and 18 on July 21, 2018.

Then, participants can submit jobs inside the container job with this (example) syntax:

$ oarsub -t inner=<job_id of the container job> -l {"network_address in ('mnode4.cortexlab.fr', 'mnode6.cortexlab.fr')"}/nodes=2,walltime=0:30:00 -I

or (another example):

$ oarsub -t inner=<job_id of the container job> -l nodes=2,walltime=0:30:00 'sleep 10000000'

OAR Documentation

The complete OAR documentation, with much more details and examples, is available here: https://oar.imag.fr

reserve.txt · Last modified: 2022/09/12 16:36 by mimbert