Incorporate CR feedback:

1. Use LinkedList rather than ArrayList for the measurement array in RAW. Update size estimation in documentation in workload_template.

2. Add a new composite measurement type which combines HdrHistogram and RAW.

3. Allow user to optionally disable output of summary stats from the raw measurement class.

core/src/main/java/com/yahoo/ycsb/measurements/Measurements.java

@@ -38,6 +38,7 @@ public enum MeasurementType {
     HISTOGRAM,
     HDRHISTOGRAM,
     HDRHISTOGRAM_AND_HISTOGRAM,
+    HDRHISTOGRAM_AND_RAW,
     TIMESERIES,
     RAW
   }
@@ -97,6 +98,10 @@ else if (mTypeString.equals("hdrhistogram+histogram"))
     {
       _measurementType = MeasurementType.HDRHISTOGRAM_AND_HISTOGRAM;
     }
+    else if (mTypeString.equals("hdrhistogram+raw"))
+    {
+      _measurementType = MeasurementType.HDRHISTOGRAM_AND_RAW;
+    }
     else if (mTypeString.equals("timeseries"))
     {
       _measurementType = MeasurementType.TIMESERIES;
@@ -139,6 +144,10 @@ OneMeasurement constructOneMeasurement(String name)
       return new TwoInOneMeasurement(name,
               new OneMeasurementHdrHistogram("Hdr"+name, _props),
               new OneMeasurementHistogram("Bucket"+name, _props));
+    case HDRHISTOGRAM_AND_RAW:
+      return new TwoInOneMeasurement(name,
+          new OneMeasurementHdrHistogram("Hdr"+name, _props),
+          new OneMeasurementHistogram("Raw"+name, _props));
     case TIMESERIES:
       return new OneMeasurementTimeSeries(name, _props);
     case RAW:

core/src/main/java/com/yahoo/ycsb/measurements/OneMeasurementRaw.java

@@ -22,7 +22,7 @@
 import java.io.IOException;
 import java.io.PrintStream;
 import java.util.Properties;
-import java.util.ArrayList;
+import java.util.LinkedList;
 import java.util.Collections;
 import java.util.Comparator;
 import com.yahoo.ycsb.measurements.exporter.MeasurementsExporter;
@@ -77,14 +77,28 @@ public int compare(RawDataPoint p1, RawDataPoint p2){
    */
   public static final String OUTPUT_FILE_PATH = "measurement.raw.output_file";
   public static final String OUTPUT_FILE_PATH_DEFAULT = "";
+
+  /**
+   * Optionally, user can request to not output summary stats. This is useful
+   * if the user chains the raw measurement type behind the HdrHistogram type
+   * which already outputs summary stats. But even in that case, the user may
+   * still want this class to compute summary stats for them, especially if
+   * they want accurate computation of percentiles (because percentils computed
+   * by histogram classes are still approximations).
+   */
+  public static final String NO_SUMMARY_STATS = "measurement.raw.no_summary";
+  public static final String NO_SUMMARY_STATS_DEFAULT = "false";
+
   private String outputFilePath = "";
   private final PrintStream outputStream;
 
-  private ArrayList<RawDataPoint> measurements;
+  private boolean noSummaryStats = false;
+
+  private LinkedList<RawDataPoint> measurements;
   private long totalLatency = 0;
 
   // A window of stats to print summary for at the next getSummary() call.
-  // It's supposed to be a one line summary, so we will just print count and 
+  // It's supposed to be a one line summary, so we will just print count and
   // average.
   private int windowOperations = 0;
   private long windowTotalLatency = 0;
@@ -111,7 +125,11 @@ public OneMeasurementRaw(String name, Properties props) {
       outputStream = System.out;
 
     }
-    measurements = new ArrayList<RawDataPoint>(1000);
+
+   noSummaryStats = Boolean.parseBoolean(props.getProperty(NO_SUMMARY_STATS,
+        NO_SUMMARY_STATS_DEFAULT));
+
+   measurements = new LinkedList<RawDataPoint>();
   }
 
   @Override
@@ -142,7 +160,7 @@ public void exportMeasurements(MeasurementsExporter exporter)
 
     int totalOps = measurements.size();
     exporter.write(getName(), "Total Operations", totalOps);
-    if (totalOps > 0) {
+    if (totalOps > 0 && !noSummaryStats) {
       exporter.write(getName(),
           "Below is a summary of latency in microseconds:", -1);
       exporter.write(getName(), "Average",

workloads/workload_template

@@ -121,11 +121,11 @@ measurementtype=histogram
 # "operation, timestamp of the measurement, latency in us"
 #
 # Raw datapoints are collected in-memory while the test is running. Each
-# data point consumes about 20 bytes (including java object overhead).
+# data point consumes about 50 bytes (including java object overhead).
 # For a typical run of 1 million to 10 million operations, this should
-# easily fit into memory. If you plan to do a run with 100s of millions of
-# operations, consider increasing your jvm heap size before you enable the
-# RAW measurement type, or split the run into multiple runs.
+# fit into memory most of the time. If you plan to do 100s of millions of
+# operations per run, consider provisioning a machine with larger RAM when using
+# the RAW measurement type, or split the run into multiple runs.
 #
 # Optionally, you can specify an output file to save raw datapoints.
 # Otherwise, raw datapoints will be written to stdout.