Interactive jobs¶

Video

Watch this in our courses: 2022 February, 2021 January

Abstract

We use the standard (and dominant) Slurm batch queuing system, all standard commands will work.
See the quick reference for the reference you need if you know Slurm or batch systems already.
Interactive jobs allow you to quickly test code (before scaling up) or getting more resources for manual analysis.
To run a single command interactively
- srun [SLURM OPTIONS] COMMAND ... to run before any COMMAND to run it in Slurm
To get an interactive shell
- srun [SLURM OPTIONS] --pty bash (general Slurm)
- sinteractive (Triton specific)

Introduction to Slurm¶

Triton is a large system that combines many different individual computer nodes. Hundreds of people are using Triton simultaneously. Thus resources (CPU time, memory, etc.) need to be shared among everyone.

This resource sharing is done by a software called a job scheduler or workload manager, and Triton’s workload manager is Slurm (which is also the dominant in the world one these days). Triton users submit jobs which are then scheduled and allocated resources by the workload manager.

An analogy: the HPC Diner

You’re eating out at the HPC Diner. What happens when you arrive?

A host greets you and takes your party size and estimated dining time.
You are given a number and asked to wait a bit.
The host looks at who is currently waiting.
If you are two people, you might squeeze in soon.
If you are a lot of people, the host will try to slowly free up enough tables to join to eat together.
If you are a really large party, you might need an advance reservation (or have to wait a really long time).
They want everyone to get a fair share of their food. Thus, people that have visited more often are asked to wait slightly longer for their table, as a balancing mechanic.

Thanks to HPC Carpentry / Sabry Razick for the idea.

Why interactive jobs?¶

There are two ways you can submit your jobs to Slurm queue system: either interactively using srun or by submitting a script using sbatch. This tutorial walks you through running your jobs interactively, and the next tutorial on serial jobs will go through serial jobs.

Some people say “the cluster is for batch computing”, but really it is to help you get your work done. Interactive jobs let you:

Run a single job in the Slurm environment to test parameters and make sure it works (which is easier than constantly modifying batch scripts).
Get a large amount of resources for some manual data analysis.

Interactive jobs¶

Let’s say you want to run the following command:

$ python3 -c 'import os; print("hi from", os.uname().nodename)'

You can submit this program to Triton using srun. All input/output still goes to your terminal (but note that graphical applications don’t work this way - see below):

$ srun --mem=100M --time=0:10:00 python3 -c 'import os; print("hi from", os.uname().nodename)'
srun: job 52204499 queued and waiting for resources

Here, we are asking for 100 Megabytes of memory (--mem=100M) for a duration of ten minutes (--time=0:10:00) (See the quick reference or below for more options). While your job - with jobid 52204499 - is waiting to be allocated resources, your shell blocks while it is waiting to continue.

You can open a new shell (ssh again) on triton and run the command squeue -u $USER or slurm q to see all the jobs you currently have waiting in queue:

$ slurm q
JOBID              PARTITION NAME                  TIME       START_TIME    STATE NODELIST(REASON)
52204499           short-ivb python3               0:00              N/A  PENDING (None)

You can see information such as the state, which partition the requested node reside in, etc.

Once resources are allocated to your job, you see the name of the machine in the Triton cluster your program ran on, output to your terminal:

srun: job 52204499 has been allocated resources
hi from ivb17.int.triton.aalto.fi

Disadvantages

Interactive jobs are useful for debugging purposes, to test your setup and configurations before you put your tasks in a batch script for later execution.

The major disadvantages include:

It blocks your shell until it finishes
If your connection to Triton gets interrupted, you lose the job and its output.

Keep in mind that you shouldn’t open 20 shells to run 20 srun jobs at once. Please have a look at the next tutorial about serial jobs.

Interactive shell¶

What if you want an actual shell to do things interactively? Put more precisely, you want access to a node in the cluster through an interactive bash shell, with many resources available, that will let you run commands such as Python and let do some basic work. For this, you just need srun’s --pty option coupled with the shell you want:

srun -p interactive --time=2:00:00 --mem=600M --pty bash

The command prompt will appear when the job starts. And you will have a bash shell runnnig on one of the computation nodes with at least 600 Megabytes of memory, for a duration of 2 hours, where you can run your programs in. The option -p interactive requests a node in the interactive partition (group of nodes) which is dedicated to interactive usage (more on this later).

Warning

Remember to exit the shell when you are done! The shell will be running if you don’t and it will count towards your usage. This wastes resources and effectively means your priority will degrade in the future.

Note

you can use sinfo to see information such as the available partitions, number of nodes in each, their time limits, etc.

Interactive shell with graphics¶

sinteractive is very similar to srun, but more clever and thus allows you to do X forwarding. It starts a screen session on the node, then sshes to there and connects to the shell:

sinteractive --time=1:00:00 --mem=1000M

Warning

Just like with srun --pty bash, remember to exit the shell. Since there is a separate screen session running, just closing the terminal isn’t enough. Exit all shells in the screen session on the node (C-d or exit) or cancel the job.

Use remote desktop if off campus

If you are off-campus, you might want to use https://vdi.aalto.fi as a virtual desktop to connect to Triton to run graphical programs: ssh from there to Triton with ssh -XY. Graphical programs run very slowly when sent across the general Internet.

Monitoring your usage¶

When your jobs enter the queue, you need to be able to get information on how much time, memory, etc. your jobs are using in order to know what requirements to ask for.

The command slurm history (or sacct --long | less) gives you information such as the actual memory used by your recent jobs, total CPU time, etc. You will learn more about these commands later on.

As shown in a previous example, the command slurm queue (or squeue -u $USER) will tell you the currently running processes, which is a good way to make sure you have stopped everything.

Note

Generally, estimating the amount of time or memory you need comes down to monitoring you Slurm history and utilizing command-line tools such as time on a few of your jobs and averaging. This is basically a trial and error process.

Setting resource parameters¶

Slurm comes with a multitude of parameters which you can specify to ensure you will be allocated enough memory, CPU cores, time, etc. You saw two of them in use in the above examples (--mem and --time) and you will learn more in the following tutorials.

Because you are sharing resource with other users, you should always estimate the amount of time, memory, etc. you need and then request them accordingly for efficiency reasons; the default memory and time limits are intentionally set low and may not be sufficient for your jobs to run/finish.

The general rule of thumb is to request the least possible, so that your stuff can run faster. That is because the less you request, the faster you are likely to be allocated resources. If you request something slightly less than a node size (note that we have different size nodes) or partition limit, you are more likely to fit into a spare spot.

For example, we have many nodes with 12 cores, and some with 20 or 24. If you request 24 cores, you have very limited options. However, you are more likely to be allocated a node if you request 10 cores. The same applies to memory: most common cutoffs are 48, 64, 128, 256GB. It’s best to use smaller values when submitting interactive jobs, and more for batch scripts.

Exercises¶

The scripts you need for the following exercises can be found in this git repository: hpc-examples. You can clone the repository by running git clone https://github.com/AaltoSciComp/hpc-examples.git. This repository will be used for most of the tutorial exercises.

Interactive-1: Basic Slurm options

The program hpc-examples/slurm/memory-hog.py uses up a lot of memory to do nothing. Let’s play with it. It’s run as follows: python hpc-examples/slurm/memory-hog.py 50M, where the last argument is however much memory you want to eat. You can use --help to see the options of the program.

Try running the program with 50M.
Run the program with 50M and srun --mem=500M.
Increase the amount of memory the Python process tries to use (not the amount of memory Slurm allocates). How much memory can you use before the job fails?
Look at the job history using slurm history - can you see how much memory it actually used? - Note that Slurm only measures memory every 60 seconds or so. To make the program last longer, so that the memory used can be measured, give the --sleep option to the Python process, like this: python hpc-examples/slurm/memory-hog.py 50M --sleep=60 - keep it available.

Interactive-2: Time scaling

The program hpc-examples/slurm/pi.py calculates pi using a simple stochastic algorithm. The program takes one positional argument: the number of trials.

The time program allows you to time any program, e.g. you can time python x.py to print the amount of time it takes.

Run the program, timing it with time, a few times, increasing the number of trials, until it takes about 10 seconds: time python hpc-examples/slurm/pi.py 500, then 5000, then 50000, and so on.
Add srun in front (srun python ...). Use the seff JOBID command to see how much time the program took to run. (If you’d like to use the time command, you can run srun --mem=MEM --time=TIME time python hpc-examples/slurm/pi.py ITERS)
Tell srun to use five CPUs (-c 5). Does it go any faster?
Use the --threads=5 option to the Python program to tell it to also use five threads. ... python .../pi.py --threads=5
Look at the job history using slurm history - can you see how much time each process used? What’s the relation between TotalCPUTime and WallTime?

Interactive-3: Info commands

Check out some of these commands: sinfo, sinfo -N, squeue. Run slurm job JOBID on some running job - does anything look interesting?

Interactive-4: Showing node information

Run scontrol show node csl1 What is this? (csl1 is the name of a node on Triton - if you are not on Triton, look at the sinfo -N command and try one of those names).

Interactive-5: Why not script srun

Some people are clever and use shell scripting to run srun many times in a loop (using & to background it so that they all run at the same time). Can you list some advantages and disadvantages to this?

What’s next?¶

In the next tutorial on serial batch jobs, you will learn how to put the above-mentioned commands in a script, namely a batch script (a.k.a submission script) that allows for a multitude of jobs to run unattended.