High Performance Computers¶
This page includes instructions and guidelines when deploying Dask on high performance supercomputers commonly found in scientific and industry research labs. These systems commonly have the following attributes:
- Some mechanism to launch MPI applications or use job schedulers like SLURM, SGE, TORQUE, LSF, DRMAA, PBS, or others
- A shared network file system visible to all machines in the cluster
- A high performance network interconnect, such as Infiniband
- Little or no node-local storage
Where to start¶
Most of this page documents best practices to use Dask on an HPC cluster. This is technical and aimed both at users with some experience deploying Dask and also system administrators.
New users may instead prefer to start with one of the following projects which provide easy high-level access to Dask using resource managers that are commonly deployed on HPC systems:
- dask-jobqueue for use with PBS, SLURM, and SGE resource managers
- dask-drmaa for use with any DRMAA compliant resource manager
They provide interfaces that look like the following:
from dask_jobqueue import PBSCluster cluster = PBSCluster(cores=36, memory="100GB", project='P48500028', queue='premium', walltime='02:00:00') cluster.start_workers(100) # Start 100 jobs that match the description above from dask.distributed import Client client = Client(cluster) # Connect to that cluster
We recommend reading the dask-jobqueue documentation first to get a basic system running and then returning to this documentation for fine-tuning.
This section is not necessary if you use a tool like dask-jobqueue.
You can launch a Dask network using
mpiexec and the
dask-mpi command line executable.
mpirun --np 4 dask-mpi --scheduler-file /home/$USER/scheduler.json
from dask.distributed import Client client = Client(scheduler_file='/path/to/scheduler.json')
This depends on the mpi4py library. It only
uses MPI to start the Dask cluster and not for inter-node communication. MPI implementations differ: the use
mpirun --np 4 is specific to the
mpich MPI implementation installed
through conda and linked to mpi4py
conda install mpi4py
It is not necessary to use exactly this implementation, but you may want to
verify that your
mpi4py Python library is linked against the proper
mpirun/mpiexec executable and that the flags used (like
--np 4) are
correct for your system. The system administrator of your cluster should be
very familiar with these concerns and able to help.
dask-mpi --help to see more options for the
High Performance Network¶
Many HPC systems have both standard Ethernet networks as well as
high-performance networks capable of increased bandwidth. You can instruct
Dask to use the high-performance network interface by using the
keyword with the
dask-mpi commands or
interface= keyword with the dask-jobqueue
mpirun --np 4 dask-mpi --scheduler-file /home/$USER/scheduler.json --interface ib0
In the code example above, we have assumed that your cluster has an Infiniband
network interface called
ib0. You can check this by asking your system
administrator or by inspecting the output of
$ 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:18.104.22.168
No Local Storage¶
Users often exceed memory limits available to a specific Dask deployment. In normal operation, Dask spills excess data to disk. However, in HPC systems, the individual compute nodes often lack locally attached storage, preferring instead to store data in a robust high performance network storage solution. As a result, when a Dask cluster starts to exceed memory limits, its workers can start making many small writes to the remote network file system. This is both inefficient (small writes to a network file system are much slower than local storage for this use case) and potentially dangerous to the file system itself.
See this page
for more information on Dask’s memory policies. Consider changing the
following values in your
distributed: worker: memory: target: false # don't spill to disk spill: false # don't spill to disk pause: 0.80 # pause execution at 80% memory use terminate: 0.95 # restart the worker at 95% use
This stops Dask workers from spilling to disk, and instead relies entirely on mechanisms to stop them from processing when they reach memory limits.
As a reminder, you can set the memory limit for a worker using the
dask-mpi ... --memory-limit 10GB
Alternatively, if you do have local storage mounted on your compute nodes, you
can point Dask workers to use a particular location in your filesystem using
dask-mpi ... --local-directory /scratch
Launch Many Small Jobs¶
HPC job schedulers are optimized for large monolithic jobs with many nodes that all need to run as a group at the same time. Dask jobs can be quite a bit more flexible: workers can come and go without strongly affecting the job. If we split our job into many smaller jobs, we can often get through the job scheduling queue much more quickly than a typical job. This is particularly valuable when we want to get started right away and interact with a Jupyter notebook session rather than waiting for hours for a suitable allocation block to become free.
So, to get a large cluster quickly, we recommend allocating a dask-scheduler process on one node with a modest wall time (the intended time of your session) and then allocating many small single-node dask-worker jobs with shorter wall times (perhaps 30 minutes) that can easily squeeze into extra space in the job scheduler. As you need more computation, you can add more of these single-node jobs or let them expire.
Use Dask to co-launch a Jupyter server¶
Dask can help you by launching other services alongside it. For example, you
can run a Jupyter notebook server on the machine running the
process with the following commands
from dask.distributed import Client client = Client(scheduler_file='scheduler.json') import socket host = client.run_on_scheduler(socket.gethostname) def start_jlab(dask_scheduler): import subprocess proc = subprocess.Popen(['/path/to/jupyter', 'lab', '--ip', host, '--no-browser']) dask_scheduler.jlab_proc = proc client.run_on_scheduler(start_jlab)