cupy-svd.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Performing SVD with Dask and CuPy\n",
    "\n",
    "This is a communication heavy computation on GPUs. Currently this is faster on GPU than CPU. Though communication improvements on the UCX should improve this further."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Use a DGX"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import dask\n",
    "import dask.config\n",
    "\n",
    "from dask.distributed import Client, wait\n",
    "from dask_cuda import LocalCUDACluster\n",
    "\n",
    "cluster = LocalCUDACluster(diagnostics_port=9100)\n",
    "client = Client(cluster)\n",
    "client"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Create Random Dataset\n",
    "\n",
    "This is done on GPU."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import dask\n",
    "import dask.array\n",
    "import cupy\n",
    "import numpy\n",
    "\n",
    "\n",
    "# CPU\n",
    "# rs = dask.array.random.RandomState(RandomState=numpy.random.RandomState)\n",
    "# GPU\n",
    "rs = dask.array.random.RandomState(RandomState=cupy.random.RandomState)\n",
    "\n",
    "x = rs.random((1000000, 1000), chunks=(10000, 1000))\n",
    "x = x.persist()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Add array with its transpose\n",
    "\n",
    "This computes SVD on GPU and has some communication heavy steps."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import dask.array.linalg\n",
    "\n",
    "u, s, v = dask.array.linalg.svd(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import time\n",
    "start = time.time()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "u, s, v = dask.persist(u, s, v)\n",
    "%time _ = wait([u, s, v])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(u.shape)\n",
    "print(s.shape)\n",
    "print(v.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "_ = client.profile(start=start, filename='dask-cupy-svd.html')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Inspect output"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(u[:10, :10].compute())\n",
    "print(s[:10].compute())\n",
    "print(v[:10, :10].compute())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from distributed.utils import format_bytes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(format_bytes(u.nbytes))\n",
    "print(format_bytes(s.nbytes))\n",
    "print(format_bytes(v.nbytes))"
   ]
  }
 ],
 "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": "3.7.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}