To properly use Dask and Dask-Jobqueue on an HPC system you need to provide a bit of information about that system and how you plan to use it.
You provide this information either as keyword arguments to the constructor:
cluster = PBSCluster(cores=36, memory='100GB', queue='regular', ...)
Or as part of a configuration file:
jobqueue: pbs: cores: 36 memory: 100GB queue: regular ...
cluster = PBSCluster()
For more information on handling configuration files see Dask configuration documentation.
This page explains what these parameters mean and how to find out information about them.
Cores and Memory¶
These numbers correspond to the size of a single job, which is typically the size of a single node on your cluster. It does not mean the total amount of cores or memory that you want for your full deployment. Recall that dask-jobqueue will launch several jobs in normal operation.
Cores should be provided as an integer, while memory is typically provided as a string, like “100 GB”.
cores: 36 memory: 100GB
By default Dask will run one Python process per job. However, you can
optionally choose to cut up that job into multiple processes using the
processes configuration value. This can be advantageous if your
computations are bound by the GIL, but disadvantageous if you plan to
communicate a lot between processes. Typically we find that for pure Numpy
workloads a low number of processes (like one) is best, while for pure Python
workloads a high number of processes (like one process per two cores) is best.
If you are unsure then you might want to experiment a bit, or just choose a
moderate number, like one process per four cores.
cores: 36 memory: 100GB processes: 9
Many HPC systems have a variety of different queues to which you can submit jobs. These typically have names like “regular”, “debug”, and “priority”. These are set up by your cluster administrators to help direct certain jobs based on their size and urgency.
If you are unfamiliar with using queues on your system you should leave this blank, or ask your IT administrator.
You may have an allocation on your HPC system that is referenced by a project. This is typically a short bit of text that references your group or a particular project. This is typically given to you by your IT administrator when they give you an allocation of hours on the HPC system.
If this sounds foreign to you or if you don’t use project codes then you should leave this blank, or ask your IT administrator.
When Dask workers run out of memory they typically start writing data to disk. This is often a wise choice on personal computers or analysis clusters, but can be unwise on HPC systems if they lack local storage. When Dask workers try to write excess data to disk on systems that lack local storage this can cause the Dask process to die in unexpected ways.
If your nodes have fast locally attached storage mounted somewhere then you should direct dask-jobqueue to use that location.
Sometimes your job scheduler will give this location to you as an environment
variable. If so you should include that environment variable, prepended with
$ sign and it will be expanded appropriately after the jobs start.
No Local Storage¶
If your nodes do not have locally attached storage then we recommend that you
turn off Dask’s policy to write excess data to disk. This must be done in a
configuration file and must be separate from the
section (though it is fine to include it in the same file).
jobqueue: pbs: cores: 36 memory: 100GB ... distributed: worker: memory: target: False # Avoid spilling to disk spill: False # Avoid spilling to disk pause: .80 # Pause worker threads at 80% use terminate: 0.95 # Restart workers at 95% use
HPC systems often have advanced networking hardware like Infiniband.
Dask workers can take use of this network using TCP-over-Infiniband, this can
yield improved bandwidth during data transfers. To get this increased speed
you often have to specify the network interface of your accelerated hardware.
If you have sufficient permissions then you can find a list of all network
interfaces using the
ifconfig UNIX command
$ ifconfig lo Link encap:Local Loopback # Localhost inet addr:127.0.0.1 Mask:255.0.0.0 inet6 addr: ::1/128 Scope:Host eth0 Link encap:Ethernet HWaddr XX:XX:XX:XX:XX:XX # Ethernet inet addr:192.168.0.101 ... ib0 Link encap:Infiniband # Fast InfiniBand inet addr:220.127.116.11
Alternatively, your IT administrators will have this information.
Managing Configuration files¶
By default when dask-jobqueue is first imported it places a file at
~/.config/dask/jobqueue.yaml with a commented out version of many different
job schedulers. You may want to do a few things to clean this up:
- Remove all of the commented out portions that don’t apply to you. For example if you use only PBS, then consider removing the entries under SGE, SLURM, etc..
- Feel free to rename the file or to include other configuration options in
the file for other parts of Dask. The
jobqueue.yamlfilename is not special, nor is it special that each component of Dask has its own configuration file. It is ok to combine or split up configuration files as suits your group.
- Ask your IT administrator to place a generic file in
/etc/daskfor global use. Dask will look first in
/etc/daskand then in
.yamlfiles preferring those in the user’s home directory to those in the
/etc/dask. By providing a global file IT should be able to provide sane settings for everyone on the same system