Dask-Jobqueue SLURMCluster with Singularity

SLURMCluster_vs_Singularity.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Dask-Jobqueue with Singularity"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Need adapter for sbatch, scancel, squeue via SSH"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "mkdir -p bin"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Overwriting bin/sbatch\n"
     ]
    }
   ],
   "source": [
    "%%file bin/sbatch\n",
    "#!/usr/bin/env bash\n",
    "\n",
    "# dirty hack: This cluster has single-tenant nodes only and hence\n",
    "# does not accept users trying to specify memory requirements.\n",
    "# So instead of teaching Dask-Jobqueue to make the memory spec optional, \n",
    "# we remove the conflicting header line just before submitting the job.\n",
    "grep -v '#SBATCH \\-\\-mem=' ${1} > ${1}.corr\n",
    "\n",
    "# SSH back to the host (var is passed through to singularity anyway).\n",
    "# Then, source /etc/profile to make it feel like an interactive session, \n",
    "# and submit the job.\n",
    "ssh $(hostname) -q -t \". /etc/profile && sbatch ${1}.corr\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Overwriting bin/scancel\n"
     ]
    }
   ],
   "source": [
    "%%file bin/scancel\n",
    "#!/usr/bin/env bash\n",
    "\n",
    "# See bin/sbatch for details\n",
    "ssh $(hostname) -q -t \". /etc/profile && scancel $@\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Overwriting bin/squeue\n"
     ]
    }
   ],
   "source": [
    "%%file bin/squeue\n",
    "#!/usr/bin/env bash\n",
    "\n",
    "# See bin/sbatch for details\n",
    "ssh $(hostname) -q -t \". /etc/profile && squeue $@\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "!chmod -R 755 bin/s*"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "pwd = !echo ${PWD}\n",
    "os.environ[\"PATH\"] = f\"{pwd[0]}/bin:{os.environ['PATH']}\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Set up dask jobqueue\n",
    "\n",
    "All we really need to do is make sure the generated job script loads the singularity module and uses a containered python executable."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "from dask_jobqueue import SLURMCluster\n",
    "from dask.distributed import Client\n",
    "import os"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "containered_python_exe = f\"singularity run {os.environ['SINGULARITY_CONTAINER']} python\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "cluster = SLURMCluster(\n",
    "    cores=40, memory=\"170G\", processes=4,\n",
    "    project=\"shk00023\",\n",
    "    queue='medium40:test', interface=\"ib0\",\n",
    "    walltime=\"00:20:00\",\n",
    "    env_extra=['module load singularity',],  # ensure singularity is loaded\n",
    "    python=containered_python_exe,  # use pyhton in container\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "#!/usr/bin/env bash\n",
      "\n",
      "#SBATCH -J dask-worker\n",
      "#SBATCH -p medium40:test\n",
      "#SBATCH -A shk00023\n",
      "#SBATCH -n 1\n",
      "#SBATCH --cpus-per-task=40\n",
      "#SBATCH --mem=159G\n",
      "#SBATCH -t 00:20:00\n",
      "module load singularity\n",
      "singularity run /home/shkifmwr/.singularity/cache/oci-tmp/0d046f7b1e98ec23133a61a71d65f550f300c8928db0d0da5ec99d4d524e962e/esm-vfc-stacks_latest.sif python -m distributed.cli.dask_worker tcp://10.246.201.2:42036 --nthreads 10 --nprocs 4 --memory-limit 42.50GB --name name --nanny --death-timeout 60 --interface ib0\n",
      "\n"
     ]
    }
   ],
   "source": [
    "print(cluster.job_script())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table style=\"border: 2px solid white;\">\n",
       "<tr>\n",
       "<td style=\"vertical-align: top; border: 0px solid white\">\n",
       "<h3 style=\"text-align: left;\">Client</h3>\n",
       "<ul style=\"text-align: left; list-style: none; margin: 0; padding: 0;\">\n",
       "  <li><b>Scheduler: </b>tcp://10.246.201.2:42036</li>\n",
       "  <li><b>Dashboard: </b><a href='http://10.246.201.2:8787/status' target='_blank'>http://10.246.201.2:8787/status</a></li>\n",
       "</ul>\n",
       "</td>\n",
       "<td style=\"vertical-align: top; border: 0px solid white\">\n",
       "<h3 style=\"text-align: left;\">Cluster</h3>\n",
       "<ul style=\"text-align: left; list-style:none; margin: 0; padding: 0;\">\n",
       "  <li><b>Workers: </b>0</li>\n",
       "  <li><b>Cores: </b>0</li>\n",
       "  <li><b>Memory: </b>0 B</li>\n",
       "</ul>\n",
       "</td>\n",
       "</tr>\n",
       "</table>"
      ],
      "text/plain": [
       "<Client: 'tcp://10.246.201.2:42036' processes=0 threads=0, memory=0 B>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "client = Client(cluster)\n",
    "display(client)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Let's scale the cluster to 12 nodes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "cluster.scale(jobs=12)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "             JOBID    PARTITION         NAME     USER  ACCOUNT     STATE       TIME NODES NODELIST(REASON)\n",
      "           2337935 medium40:tes  dask-worker shkifmwr shk00023   PENDING       0:00     1 (None)\n",
      "           2337931 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:01     1 gcn1276\n",
      "           2337932 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:01     1 gcn1277\n",
      "           2337933 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:01     1 gcn1033\n",
      "           2337934 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:01     1 gcn1034\n",
      "           2337928 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:04     1 gcn1108\n",
      "           2337929 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:04     1 gcn1224\n",
      "           2337930 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:04     1 gcn1225\n",
      "           2337925 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:07     1 gcn1353\n",
      "           2337926 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:07     1 gcn1404\n",
      "           2337927 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:07     1 gcn1107\n",
      "           2337924 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       0:10     1 gcn1141\n"
     ]
    }
   ],
   "source": [
    "!bin/squeue -u $USER"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table style=\"border: 2px solid white;\">\n",
       "<tr>\n",
       "<td style=\"vertical-align: top; border: 0px solid white\">\n",
       "<h3 style=\"text-align: left;\">Client</h3>\n",
       "<ul style=\"text-align: left; list-style: none; margin: 0; padding: 0;\">\n",
       "  <li><b>Scheduler: </b>tcp://10.246.201.2:42036</li>\n",
       "  <li><b>Dashboard: </b><a href='http://10.246.201.2:8787/status' target='_blank'>http://10.246.201.2:8787/status</a></li>\n",
       "</ul>\n",
       "</td>\n",
       "<td style=\"vertical-align: top; border: 0px solid white\">\n",
       "<h3 style=\"text-align: left;\">Cluster</h3>\n",
       "<ul style=\"text-align: left; list-style:none; margin: 0; padding: 0;\">\n",
       "  <li><b>Workers: </b>40</li>\n",
       "  <li><b>Cores: </b>400</li>\n",
       "  <li><b>Memory: </b>1.70 TB</li>\n",
       "</ul>\n",
       "</td>\n",
       "</tr>\n",
       "</table>"
      ],
      "text/plain": [
       "<Client: 'tcp://10.246.201.2:42036' processes=28 threads=280, memory=1.19 TB>"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "client"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## A calculation\n",
    "\n",
    "We use my [goto example for Dask array demos](https://nbviewer.jupyter.org/github/willirath/dask_jobqueue_workshop_materials/blob/15874fe653f2c129fd47b3a156d9a16292978505/notebooks/01_local_cluster_monte_carlo_estimate_of_pi.ipynb): An embarassingly parallel Monte Carlo estimate of the number Pi.\n",
    "\n",
    "[<img src=\"https://upload.wikimedia.org/wikipedia/commons/8/84/Pi_30K.gif\" width=\"300px\" alt=\"PI monte-carlo estimate\">](https://en.wikipedia.org/wiki/Pi#Monte_Carlo_methods)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "from dask import array as darr\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "from 1000.0 GB of random data:\tpi = 3.141594476352, err = 1.8227622069488802e-06\n",
      "\n",
      "CPU times: user 7.31 s, sys: 656 ms, total: 7.97 s\n",
      "Wall time: 11.3 s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "xy = darr.random.uniform(\n",
    "    0, 1,\n",
    "    size=(1e12 / 16, 2), chunks=(500e6 / 16, None)\n",
    ")\n",
    "pi = 4 * ((xy ** 2).sum(-1) < 1).mean()\n",
    "\n",
    "pi = pi.compute()\n",
    "\n",
    "print(f\"from {xy.nbytes / 1e9} GB of random data:\\tpi = {pi}, err = {pi - np.pi}\\n\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Adaptive clusters\n",
    "\n",
    "We'll set a target duration of 2 minutes to perform the same calculation with 5 TB of random data."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "ca = cluster.adapt(\n",
    "    minimum=4, maximum=48,\n",
    "    target_duration=\"1200\",  # measured in CPU time per worker\n",
    "                             # -> 120 seconds at 10 cores / worker\n",
    "    wait_count=4  # scale down more gently\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "             JOBID    PARTITION         NAME     USER  ACCOUNT     STATE       TIME NODES NODELIST(REASON)\n",
      "           2337927 medium40:tes  dask-worker shkifmwr shk00023   RUNNING       1:26     1 gcn1107\n"
     ]
    }
   ],
   "source": [
    "!bin/squeue -u $USER"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "from 5000.0 GB of random data:\tpi = 3.1415927031936, err = 4.960380683982635e-08\n",
      "\n",
      "CPU times: user 1min 19s, sys: 8.31 s, total: 1min 27s\n",
      "Wall time: 2min 53s\n"
     ]
    }
   ],
   "source": [
    "%%time\n",
    "\n",
    "xy = darr.random.uniform(\n",
    "    0, 1,\n",
    "    size=(5e12 / 16, 2), chunks=(500e6 / 16, None)\n",
    ")\n",
    "pi = 4 * ((xy ** 2).sum(-1) < 1).mean()\n",
    "\n",
    "pi = pi.compute()\n",
    "\n",
    "print(f\"from {xy.nbytes / 1e9} GB of random data:\\tpi = {pi}, err = {pi - np.pi}\\n\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}