Advanced tips and tricks

The universe of HPC clusters is extremely diverse, with different job schedulers, different configuration, different decisions (security, usage, etc…) made by each HPC cluster. An unfortunate consequence of this is that this is impossible for Dask-Jobqueue to cover all possible tiny edge cases of some HPC clusters.

This page is an attempt to document tips and tricks that are likely to be useful on some clusters (strictly more than one ideally although hard to be sure …).

Skipping unrecognised line in submission script with job_directives_skip

Note: the parameter job_directives_skip was named header_skip until version 0.8.0. header_skip can still be used, but is considered deprecated and will be removed in a future version.

On some clusters, the submission script generated by Dask-Jobqueue (you can look at it with print(cluster.job_script())) may not work because on some configuration quirk of this HPC cluster. Probably there are some reasons behind this configuration quirk of course.

You’ll get an error when calling cluster.scale (i.e. where you actually submit some jobs) that will tell you the job scheduler is not happy with your job submission script (see examples below). The main parameter you can use to work-around this is job_directives_skip:

# this will remove any line containing either '--mem' or
# 'another-string' from the job submission script
cluster = YourCluster(
    job_directives_skip=['--mem', 'another-string'],
    **other_options_go_here)

An example of this problem is very well detailed in this blog post by Matthew Rocklin. In his case, the error was:

Command:
bsub /tmp/tmp4874eufw.sh
stdout:

Typical usage:
     bsub [LSF arguments] jobscript
     bsub [LSF arguments] -Is $SHELL
     bsub -h[elp] [options]
     bsub -V

NOTES:
 * All jobs must specify a walltime (-W) and project id (-P)
 * Standard jobs must specify a node count (-nnodes) or -ln_slots. These jobs cannot specify a resource string (-R).
 * Expert mode jobs (-csm y) must specify a resource string and cannot specify -nnodes or -ln_slots.

stderr:
ERROR: Resource strings (-R) are not supported in easy mode. Please resubmit without a resource string.
ERROR: -n is no longer supported. Please request nodes with -nnodes.
ERROR: No nodes requested. Please request nodes with -nnodes.

Another example of this issue is this github issue where --mem is not an accepted option on some SLURM clusters. The error was something like this:

$sbatch submit_slurm.sh
sbatch: error: Memory specification can not be satisfied
sbatch: error: Batch job submission failed: Requested node configuration is not available

Run setup commands before starting the worker with job_script_prologue

Note: the parameter job_script_prologue was named env_extra until version 0.7.4. env_extra can still be used, but is considered deprecated and will be removed in a future version.

Sometimes you need to run some setup commands before the actual worker can be started. This includes setting environment variables, loading environment modules, sourcing/activating a virtual environment, or activating conda/mamba environments.

This can be achieved using the job_script_prologue parameter. Example for setting up a virtual environment:

from dask_jobqueue.htcondor import HTCondorCluster
job_script_prologue = ['cd /some/path', 'source venv/bin/activate']
cluster = HTCondorCluster(cores=1, memory="2GB", disk="4GB", log_directory = 'logs', python='python3',
                          job_script_prologue=job_script_prologue)
print(cluster.job_script())

For HTCondorCluster, the commands will be prepended to the actual python call in the Arguments parameter in the submit description file. The relevant lines will look like this:

...
Arguments = "-c 'cd /some/path; source venv/bin/activate; python3 -m distributed.cli.dask_worker tcp://<IP>:<PORT> --nthreads 1 --memory-limit 2.00GB --name dummy-name --nanny --death-timeout 60'"
Executable = /bin/sh
...

For other batch systems (*Cluster classes) the additional commands will be inserted as separate lines in the submission script.

How to handle job queueing system walltime killing workers

In dask-jobqueue, every worker process runs inside a job, and all jobs have a time limit in job queueing systems. Reaching walltime can be troublesome in several cases:

  • when you don’t have a lot of room on you HPC platform and have only a few workers at a time (less than what you were hoping for when using scale or adapt). These workers will be killed (and others started) before your workload ends.

  • when you really don’t know how long your workload will take: all your workers could be killed before reaching the end. In this case, you’ll want to use adaptive clusters so that Dask ensures some workers are always up.

If you don’t set the proper parameters, you’ll run into KilleWorker exception in those two cases.

The solution to this problem is to tell Dask up front that the workers have a finite lifetime:

  • Use –lifetime worker option. This will enable infinite workloads using adaptive. Workers will be properly shut down before the scheduling system kills them, and all their states moved.

  • Use –lifetime-stagger when dealing with many workers (say > 20): this will prevent workers from terminating at the same time, thus ease rebalancing tasks and scheduling burden.

Here is an example of how to use these parameters:

cluster = Cluster(
    walltime="01:00:00",
    cores=4,
    memory="16gb",
    worker_extra_args=["--lifetime", "55m", "--lifetime-stagger", "4m"],
)
cluster.adapt(minimum=0, maximum=200)

Note: the parameter worker_extra_args was named extra until version 0.7.4. extra can still be used, but is considered deprecated and will be removed in a future version.

Here is an example of a workflow taking advantage of this, if you want to give it a try or adapt it to your use case:

import time
import numpy as np
from dask_jobqueue import PBSCluster as Cluster
from dask import delayed
from dask.distributed import Client, as_completed

# config in $HOME/.config/dask/jobqueue.yaml
cluster = Cluster(walltime='00:01:00', cores=1, memory='4gb')
cluster.adapt(minimum=0, maximum=4)

client = Client(cluster)

# each job takes 1s, and we have 4 cpus * 1 min * 60s/min = 240 cpu.s, let's ask for a little more tasks.
filenames = [f'img{num}.jpg' for num in range(480)]

def features(num_fn):
    num, image_fn = num_fn
    time.sleep(1)  # takes about 1s to compute features on an image
    features = np.random.random(246)
    return num, features

num_files = len(filenames)
num_features = len(features((0, filenames[0]))[1]) # FIX

X = np.zeros((num_files, num_features), dtype=np.float32)

for future in as_completed(client.map(features, list(enumerate(filenames)))): # FIX
    i, v = future.result()
    X[i, :] = v