Dask

From Alliance Doc
Jump to navigation Jump to search
This site replaces the former Compute Canada documentation site, and is now being managed by the Digital Research Alliance of Canada.

Ce site remplace l'ancien site de documentation de Calcul Canada et est maintenant géré par l'Alliance de recherche numérique du Canada.

Dask is a flexible library for parallel computing in Python. It provides parallelized NumPy array and Pandas DataFrame objects, and it enables distributed computing in pure Python with access to the PyData stack.

Installing our wheel

The preferred option is to install it using our provided Python wheel as follows:

1. Load a Python module, thus module load python/3.11
2. Create and start a virtual environment.
3. Install dask, and optionally dask-distributed in the virtual environment with pip install.
Question.png
(venv) [name@server ~] pip install --no-index dask distributed

Job Submission

Single Node

Below is an example of a job that spawns a single-node Dask cluster with 6 cpus and computes the mean of a column of a parallelized dataframe.


File : dask-example.sh

#!/bin/bash
#SBATCH --account=<your account>
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=6  
#SBATCH --mem=8000M       
#SBATCH --time=0-00:05
#SBATCH --output=%N-%j.out

module load python/3.11
virtualenv --no-download $SLURM_TMPDIR/env
source $SLURM_TMPDIR/env/bin/activate

pip install dask distributed pandas --no-index

python dask-example.py


In the script Dask-example.py, we launch a Dask cluster with as many worker processes as there are cores in our job. This means each worker will spawn at most one CPU thread. For a complete discussion of how to reason about the number worker processes and the number of threads per worker, see the official Dask documentation. In this example, we split a pandas data frame into 6 chunks, so each worker will process a part of the data frame using one CPU:


File : dask-example.py

import pandas as pd

from dask import dataframe as dd
from dask.distributed import Client, LocalCluster

import os

n_workers = os.environ['SLURM_CPUS_PER_TASK']

cluster = LocalCluster(n_workers=n_workers, threads_per_worker=1)
client = Client(cluster)

index = pd.date_range("2021-09-01", periods=2400, freq="1H")
df = pd.DataFrame({"a": np.arange(2400)}, index=index)
ddf = dd.from_pandas(df, npartitions=n_workers) # split the pandas data frame into "n_workers" chunks

result = ddf.a.mean().compute()

print(f"The mean is {result}")


Multiple Nodes

In the example that follows, we reproduce the single-node example, but this time with a two-node Dask cluster, with 6 CPUs on each node. This time we also spawn 2 workers per node, each with 3 cores.


File : dask-example.sh

#!/bin/bash
#SBATCH --nodes 2
#SBATCH --tasks-per-node=2
#SBATCH --mem=16000M
#SBATCH --cpus-per-task=3
#SBATCH --time=0-00:30
#SBATCH --output=%N-%j.out
#SBATCH --account=<your account>

export DASK_SCHEDULER_ADDR=$(hostname)
export DASK_SCHEDULER_PORT=34567

srun -N 2 -n 2 config_virtualenv.sh # set both -N and -n to the number of nodes

source $SLURM_TMPDIR/ENV/bin/activate

dask scheduler --host $DASK_SCHEDULER_ADDR --port $DASK_SCHEDULER_PORT &
sleep 10

srun launch_dask_workers.sh &
dask_cluster_pid=$!
sleep 10

python test_dask.py

kill $dask_cluster_pid # shut down Dask workers after the python process exits


Where the script config_env.sh is:


File : config_env.sh

#!/bin/bash

echo "From node ${SLURM_NODEID}: installing virtualenv..."

module load python/3.11
virtualenv --no-download $SLURM_TMPDIR/env
source $SLURM_TMPDIR/env/bin/activate

pip install --no-index dask distributed pandas

echo "Done installing virtualenv!"

deactivate


And the script launch_dask_workers.sh is:


File : launch_dask_workers.sh

#!/bin/bash

source $SLURM_TMPDIR/env/bin/activate

SCHEDULER_CONNECTION_STRING="tcp://$DASK_SCHEDULER_ADDR:$DASK_SCHEDULER_PORT"

if [[ "$SLURM_PROCID" -eq "0" ]]; then
## On the SLURM task with Rank 0, where the Dask scheduler process has already been launched, we launch a smaller worker,
## with 40% of the job's memory and we subtract one core from the task to leave it for the scheduler.
        DASK_WORKER_MEM=0.4
        DASK_WORKER_THREADS=$(($SLURM_CPUS_PER_TASK-1))

else
## On all other SLURM tasks, each worker gets half of the job's allocated memory and all the cores allocated to its task.
        DASK_WORKER_MEM=0.5
        DASK_WORKER_THREADS=$SLURM_CPUS_PER_TASK
fi

dask worker "tcp://$DASK_SCHEDULER_ADDR:$DASK_SCHEDULER_PORT" --no-dashboard --nworkers=1 \
--nthreads=$DASK_WORKER_THREADS --memory-limit=$DASK_WORKER_MEM --local-directory=$SLURM_TMPDIR

sleep 5
echo "dask worker started!"


And, finally, the script test_dask.py is:


File : test_dask.py

import pandas as pd

from dask import dataframe as dd
from dask.distributed import Client

import os

client = Client(f"tcp://{os.environ['DASK_SCHEDULER_ADDR']}:{os.environ['DASK_SCHEDULER_PORT']}")

index = pd.date_range("2021-09-01", periods=2400, freq="1H")
df = pd.DataFrame({"a": np.arange(2400)}, index=index)
ddf = dd.from_pandas(df, npartitions=6)

result = ddf.a.mean().compute()

print(f"The mean is {result}")