feat: use commons-math3 for standard deviation

metacosm · Jul 7, 2024 · 4cb0e19 · 4cb0e19
1 parent f0a9de5
commit 4cb0e19
Show file tree

Hide file tree

Showing 3 changed files with 59 additions and 59 deletions.
diff --git a/measure/pom.xml b/measure/pom.xml
@@ -18,6 +18,11 @@
       <artifactId>power-server-metadata</artifactId>
       <version>${project.version}</version>
     </dependency>
+    <dependency>
+      <groupId>org.apache.commons</groupId>
+      <artifactId>commons-math3</artifactId>
+      <version>3.6.1</version>
+    </dependency>
   </dependencies>
 
   <build>

diff --git a/measure/src/main/java/net/laprun/sustainability/power/measure/PowerMeasure.java b/measure/src/main/java/net/laprun/sustainability/power/measure/PowerMeasure.java
@@ -1,31 +1,14 @@
 package net.laprun.sustainability.power.measure;
 
 import java.time.Duration;
-import java.util.Arrays;
 import java.util.List;
 import java.util.stream.IntStream;
 
+import org.apache.commons.math3.stat.descriptive.DescriptiveStatistics;
+
 import net.laprun.sustainability.power.SensorMetadata;
 
 public interface PowerMeasure {
-    int numberOfSamples();
-
-    Duration duration();
-
-    default double average() {
-        return total() / numberOfSamples();
-    }
-
-    double total();
-
-    SensorMetadata metadata();
-
-    double[] averagesPerComponent();
-
-    double minMeasuredTotal();
-
-    double maxMeasuredTotal();
-
     private static double sumOfComponents(double[] recorded) {
         var componentSum = 0.0;
         for (double value : recorded) {
@@ -45,35 +28,65 @@ static double sumOfSelectedComponents(double[] recorded, int... indices) {
         return componentSum;
     }
 
+    static String asString(PowerMeasure measure) {
+        final var durationInSeconds = measure.duration().getSeconds();
+        final var samples = measure.numberOfSamples();
+        final var measuredMilliWatts = measure.total();
+        final var stdDevs = measure.standardDeviations();
+        return String.format("%s / avg: %s / std dev: %.3f [min: %.3f, max: %.3f] (%ds, %s samples)",
+                readableWithUnit(measuredMilliWatts), readableWithUnit(measure.average()), stdDevs.aggregate,
+                measure.minMeasuredTotal(), measure.maxMeasuredTotal(), durationInSeconds, samples);
+    }
+
+    static String readableWithUnit(double milliWatts) {
+        String unit = milliWatts >= 1000 ? "W" : "mW";
+        double power = milliWatts >= 1000 ? milliWatts / 1000 : milliWatts;
+        return String.format("%.3f%s", power, unit);
+    }
+
+    int numberOfSamples();
+
+    Duration duration();
+
+    default double average() {
+        return total() / numberOfSamples();
+    }
+
+    double total();
+
+    SensorMetadata metadata();
+
+    double[] averagesPerComponent();
+
+    double minMeasuredTotal();
+
+    double maxMeasuredTotal();
+
     default StdDev standardDeviations() {
         final var cardinality = metadata().componentCardinality();
         final var totalComponents = metadata().totalComponents();
-        final var stdDevs = new double[cardinality];
-        final var aggregate = new double[1];
-        final var samples = numberOfSamples() - 1; // unbiased so we remove one sample
-        final var sqrdAverages = Arrays.stream(averagesPerComponent()).map(m -> m * m).toArray();
-        final var sqrdAverage = average() * average();
-        // need to compute the average of variances then square root that to get the "aggregate" standard deviation,
-        // see: https://stats.stackexchange.com/a/26647
-        // "vectorize" computation of variances: compute the variance for each component in parallel
+        final DescriptiveStatistics[] perComponent = new DescriptiveStatistics[cardinality];
+        for (int i = 0; i < perComponent.length; i++) {
+            perComponent[i] = new DescriptiveStatistics();
+        }
+        final DescriptiveStatistics total = new DescriptiveStatistics();
         IntStream.range(0, cardinality).parallel()
-                // compute variances for each component of the measure
                 .forEach(component -> {
-                    final var sumOfSquares = measures().stream().parallel().peek(m -> {
-                        // compute the std dev for total measure
-                        final var total = sumOfSelectedComponents(m, totalComponents);
-                        aggregate[0] += total * total;
-                    }).mapToDouble(m -> m[component] * m[component]).sum();
-                    stdDevs[component] = stdDev(sumOfSquares, sqrdAverages[component], samples);
-                    aggregate[0] = stdDev(aggregate[0], sqrdAverage, samples);
+                    measures().stream().parallel().forEach(measure -> {
+                        perComponent[component].addValue(measure[component]);
+                        total.addValue(sumOfSelectedComponents(measure, totalComponents));
+                    });
                 });
-        return new StdDev(aggregate[0], stdDevs);
-    }
 
-    private static double stdDev(double sumOfSquares, double squaredAvg, int samples) {
-        return Math.sqrt((sumOfSquares / samples) - (((samples + 1) * squaredAvg) / samples));
+        final var stdDevs = new double[cardinality];
+        for (int i = 0; i < perComponent.length; i++) {
+            stdDevs[i] = perComponent[i].getStandardDeviation();
+        }
+        return new StdDev(total.getStandardDeviation(), stdDevs);
     }
 
+    List<double[]> measures();
+
     /**
      * Records the standard deviations for the aggregated energy comsumption value (as returned by {@link #total()}) and
      * per component
@@ -83,22 +96,4 @@ private static double stdDev(double sumOfSquares, double squaredAvg, int samples
      */
     record StdDev(double aggregate, double[] perComponent) {
     }
-
-    static String asString(PowerMeasure measure) {
-        final var durationInSeconds = measure.duration().getSeconds();
-        final var samples = measure.numberOfSamples();
-        final var measuredMilliWatts = measure.total();
-        final var stdDevs = measure.standardDeviations();
-        return String.format("%s / avg: %s / std dev: %.3f [min: %.3f, max: %.3f] (%ds, %s samples)",
-                readableWithUnit(measuredMilliWatts), readableWithUnit(measure.average()), stdDevs.aggregate,
-                measure.minMeasuredTotal(), measure.maxMeasuredTotal(), durationInSeconds, samples);
-    }
-
-    static String readableWithUnit(double milliWatts) {
-        String unit = milliWatts >= 1000 ? "W" : "mW";
-        double power = milliWatts >= 1000 ? milliWatts / 1000 : milliWatts;
-        return String.format("%.3f%s", power, unit);
-    }
-
-    List<double[]> measures();
 }
diff --git a/measure/src/test/java/net/laprun/sustainability/power/measure/OngoingPowerMeasureTest.java b/measure/src/test/java/net/laprun/sustainability/power/measure/OngoingPowerMeasureTest.java
@@ -43,8 +43,8 @@ public int componentCardinality() {
         assertEquals((m1c1 + m2c1) / 2, c1Avg);
         assertEquals((m1c2 + m2c2) / 2, c2Avg);
 
-        final var stdVarForC1 = Math.sqrt(Math.pow(m1c1 - c1Avg, 2) + Math.pow(m2c1 - c1Avg, 2));
-        final var stdVarForC2 = Math.sqrt(Math.pow(m1c2 - c2Avg, 2) + Math.pow(m2c2 - c2Avg, 2));
+        final var stdVarForC1 = Math.sqrt((Math.pow(m1c1 - c1Avg, 2) + Math.pow(m2c1 - c1Avg, 2)) / (2 - 1));
+        final var stdVarForC2 = Math.sqrt((Math.pow(m1c2 - c2Avg, 2) + Math.pow(m2c2 - c2Avg, 2)) / (2 - 1));
 
         assertEquals(stdVarForC1, measure.standardDeviations().perComponent()[0], 0.0001,
                 "Standard Deviation did not match the expected value");