Why is this LR code too slow?

Question

because MLlib does not support sparse input. Therefore, I run the current code, which supports a sparse input format, on spark clusters. And setting:

• 5 nodes, each node with 8 cores (all processors on each node are 100%, 98% for the user model, when you run the code).
• input: instance 10,000,000+ and 600,000+ size on HDFS

Code:

import java.util.Random
import scala.collection.mutable.HashMap
import scala.io.Source
import org.apache.spark.SparkContext
import org.apache.spark.rdd.RDD
import org.apache.spark.util.Vector
import java.lang.Math
import org.apache.spark.broadcast.Broadcast

object SparseLR {
  val lableNum = 1
  val dimNum = 632918 
  val iteration = 10
  val alpha = 0.1
  val lambda = 0.1
  val rand = new Random(42)
  var w = Vector(dimNum, _=> rand.nextDouble)

  class SparserVector {
    var elements = new HashMap[Int, Double]

    def insert(index: Int, value: Double){
      elements += index -> value;
    }


    def *(scale: Double): Vector = {
      var x = new Array[Double](dimNum)
      elements.keySet.foreach(k => x(k) = scale * elements.get(k).get)
      Vector(x)
    }
  }
  case class DataPoint(x: SparserVector, y: Int)

  def parsePoint(line: String): DataPoint = {
    var features = new SparserVector
    val fields = line.split("\t")
    //println("fields:" + fields(0))
    val y = fields(0).toInt
    fields.filter(_.contains(":")).foreach( f => {
      val feature = f.split(":")
      features.insert(feature(0).toInt, feature(1).toDouble)
    })
    return DataPoint(features, y)
  }

  def gradient(p: DataPoint, w: Broadcast[Vector]) : Vector = {
    def h(w: Broadcast[Vector], x: SparserVector): Double = {
      val wb = w.value
      val features = x.elements
      val s = features.keySet.map(k => features.get(k).get * wb(k)).reduce(_ + _)
      1 / (1 + Math.exp(-p.y * s))
    }
    p.x * (-(1 - p.y *h(w, p.x)))
  }

  def train(sc: SparkContext, dataPoints: RDD[DataPoint]) {
    //val sampleNum = dataPoints.count
    val sampleNum = 11680250

    for(i <- 0 until iteration) {
      val wb = sc.broadcast(w)
      val g = (dataPoints.map(p => gradient(p, wb)).reduce(_ + _) + lambda * wb.value) /sampleNum
      w -= alpha * g

      println("iteration " + i + ": g = " + g)
    }
  }

  def main(args : Array[String]): Unit = {
    System.setProperty("spark.executor.memory", "15g")
    System.setProperty("spark.default.parallelism", "32");
    val sc = new SparkContext("spark://xxx:12036", "LR", "/xxx/spark", List("xxx_2.9.3-1.0.jar"))
    val lines = sc.textFile("hdfs:xxx/xxx.txt", 32)

    val trainset = lines.map(parsePoint _).cache()

    train(sc, trainset)
  }
}

Can anyone help me? Thanks!

Answer 1

In fact, it is difficult for you to give an answer. Perhaps this would be a better match for viewing the stackoverflow subdirectory code ?

, :

. - / ,

for((k,v)<-map) { 
  ...
}

for(k<-map.keySet) { val value = map.get(k).get; 
  ... 
}

, . ,

def gradient(p: DataPoint, w: Broadcast[Vector]) : Vector = {
  def h(w: Broadcast[Vector], x: SparserVector): Double = {
    val wb = w.value
    val features = x.elements
    var s = 0.0
    for((k,v)<-features)
      s += v * wb(k)
    1 / (1 + Math.exp(-p.y * s))
  }
  p.x * (-(1 - p.y *h(w, p.x)))
}

, , SparseVector, Map [Int, Double]. , , Array [Int] Array [Double] -

( , SortedMap [Int, Double] )

class SparseVector(val indices: Array[Int], val values: Array[Double]) {
  require(indices.length == values.length)

  def *(scale: Double): Vector = {
    var x = new Array[Double](dimNum)
    var i = 0
    while(i < indices.length) {  
      x(indices(i)) = scale * values(i) 
      i += 1
    }
    Vector(x)
  }
}

, , , .