SPARK-10478_microbenchmarks.md

[SPARK-10478] SoftmaxFunction.eval Benchmarks Before/After PR 8648

Num Examples	Num Classes	Before Average Runtime (ms)	After Average Runtime (ms)
1000	5	5.3315174999999995	0.4096546
1000	50	6.7308793	4.1289975
1000	250	21.5866622	21.25164
10000	5	4.2153461	4.145396
10000	50	42.2423785	41.0198213
10000	250	215.2192236	208.61861249999998
100000	5	43.517224899999995	41.2843315
100000	50	439.6628608	441.096843
100000	250	2158.0049358	2518.1192571999995

To Reproduce:

Remove private[ann] modifier to ml.ann.SoftmaxFunction, compile, paste into spark-shell.

First, paste in this:

import breeze.linalg.{DenseMatrix => BDM}
import org.apache.log4j._

import org.apache.spark.ml.ann.SoftmaxFunction

Logger.getRootLogger.setLevel(Level.OFF)

val numTrials = 10

val softmax = new SoftmaxFunction()

Then this:

for {
  numExamples <- Seq(1000, 10000, 100000);
  numClasses <- Seq(5, 50, 250)
} {
  val times = Seq.fill(numTrials) {
    val x = BDM.rand[Double](numExamples, numClasses)
    val y = BDM.zeros[Double](numExamples, numClasses)

    val startTime = System.nanoTime()

    softmax.eval(x,y)

    val stopTime = System.nanoTime()

    stopTime - startTime
  }
  val avgTime = (times.reduce(_ + _).toDouble / times.size) * 1e-6

  println(s"numExamples: $numExamples, numClasses: $numClasses, avgTime: $avgTime ms")
}