honeymoose
/
OpenSearch
mirror of https://github.com/honeymoose/OpenSearch.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

HistogramAggregator: Finer-grained rounding. 

The way `HistogramAggregator` works is that for every value, it is going to
compute a rounded value, that basically looks like
`(value / interval) * interval` and use it as a key in a hash table to
aggregate counts.

However, the exact rounded value is not needed yet at that stage, all we need
is a value that uniquely identifies the bucket, such as `(value / interval)`.
We could only multiply with `interval` again when building the bucket: this way
the second step is only performed once per bucket instead of once per value.

Although this looks like a micro optimization for the case that was just
decribed, it makes more sense with the date rounding implementations that we
have that are more CPU-intensive.

Close #4800
This commit is contained in:
Adrien Grand 2014-01-19 23:50:20 +01:00
parent 92a026b3b9
commit 1047267021
3 changed files with 90 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
src/main/java/org/elasticsearch/common/rounding/Rounding.java
View File

@ -28,17 +28,29 @@ import java.io.IOException;
/**
* A strategy for rounding long values.
*/
public interface Rounding extends Streamable {
public abstract class Rounding implements Streamable {
byte id();
public abstract byte id();
/**
* Rounds the given value.
* Given a value, compute a key that uniquely identifies the rounded value although it is not necessarily equal to the rounding value itself.
*/
public abstract long roundKey(long value);
/**
* Compute the rounded value given the key that identifies it.
*/
public abstract long valueForKey(long key);
/**
* Rounds the given value, equivalent to calling <code>roundValue(roundKey(value))</code>.
*
* @param value The value to round.
* @return The rounded value.
*/
long round(long value);
public final long round(long value) {
return valueForKey(roundKey(value));
}
/**
* Given the rounded value (which was potentially generated by {@link #round(long)}, returns the next rounding value. For example, with
@ -47,14 +59,14 @@ public interface Rounding extends Streamable {
* @param value The current rounding value
* @return The next rounding value;
*/
long nextRoundingValue(long value);
public abstract long nextRoundingValue(long value);
/**
* Rounding strategy which is based on an interval
*
* {@code rounded = value - (value % interval) }
*/
public static class Interval implements Rounding {
public static class Interval extends Rounding {
final static byte ID = 0;
@ -77,21 +89,26 @@ public interface Rounding extends Streamable {
return ID;
}
static long round(long value, long interval) {
long rem = value % interval;
// We need this condition because % may return a negative result on negative numbers
// According to Google caliper's IntModBenchmark, using a condition is faster than attempts to use tricks to avoid
// the condition. Moreover, in our case, the condition is very likely to be always true (dates, prices, distances),
// so easily predictable by the CPU
if (rem < 0) {
rem += interval;
public static long roundKey(long value, long interval) {
if (value < 0) {
return (value - interval + 1) / interval;
} else {
return value / interval;
}
return value - rem;
}
public static long roundValue(long key, long interval) {
return key * interval;
}
@Override
public long round(long value) {
return round(value, interval);
public long roundKey(long value) {
return roundKey(value, interval);
}
@Override
public long valueForKey(long key) {
return key * interval;
}
@Override

73
src/main/java/org/elasticsearch/common/rounding/TimeZoneRounding.java
View File

@ -29,9 +29,7 @@ import java.io.IOException;
/**
*/
public abstract class TimeZoneRounding implements Rounding {
public abstract long round(long utcMillis);
public abstract class TimeZoneRounding extends Rounding {
public static Builder builder(DateTimeUnit unit) {
return new Builder(unit);
@ -148,9 +146,13 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
public long roundKey(long utcMillis) {
long time = utcMillis + preTz.getOffset(utcMillis);
time = unit.field().roundFloor(time);
return unit.field().roundFloor(time);
}
@Override
public long valueForKey(long time) {
// now, time is still in local, move it to UTC (or the adjustLargeInterval flag is set)
time = time - preTz.getOffset(time);
// now apply post Tz
@ -160,7 +162,6 @@ public abstract class TimeZoneRounding implements Rounding {
@Override
public long nextRoundingValue(long value) {
// return value + unit.field().getDurationField().getUnitMillis();
return unit.field().roundCeiling(value + 1);
}
@ -198,10 +199,15 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
public long roundKey(long utcMillis) {
return unit.field().roundFloor(utcMillis);
}
@Override
public long valueForKey(long key) {
return key;
}
@Override
public long nextRoundingValue(long value) {
return unit.field().roundCeiling(value + 1);
@ -241,9 +247,13 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
public long roundKey(long utcMillis) {
long time = utcMillis + preTz.getOffset(utcMillis);
time = unit.field().roundFloor(time);
return unit.field().roundFloor(time);
}
@Override
public long valueForKey(long time) {
// after rounding, since its day level (and above), its actually UTC!
// now apply post Tz
time = time + postTz.getOffset(time);
@ -289,8 +299,13 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
return Rounding.Interval.round(utcMillis, interval);
public long roundKey(long utcMillis) {
return Rounding.Interval.roundKey(utcMillis, interval);
}
@Override
public long valueForKey(long key) {
return Rounding.Interval.roundValue(key, interval);
}
@Override
@ -333,9 +348,14 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
public long roundKey(long utcMillis) {
long time = utcMillis + preTz.getOffset(utcMillis);
time = Rounding.Interval.round(time, interval);
return Rounding.Interval.roundKey(time, interval);
}
@Override
public long valueForKey(long key) {
long time = Rounding.Interval.roundValue(key, interval);
// now, time is still in local, move it to UTC
time = time - preTz.getOffset(time);
// now apply post Tz
@ -386,9 +406,14 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
public long roundKey(long utcMillis) {
long time = utcMillis + preTz.getOffset(utcMillis);
time = Rounding.Interval.round(time, interval);
return Rounding.Interval.roundKey(time, interval);
}
@Override
public long valueForKey(long key) {
long time = Rounding.Interval.roundValue(key, interval);
// after rounding, since its day level (and above), its actually UTC!
// now apply post Tz
time = time + postTz.getOffset(time);
@ -437,8 +462,13 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
return timeZoneRounding.round((long) (factor * utcMillis));
public long roundKey(long utcMillis) {
return timeZoneRounding.roundKey((long) (factor * utcMillis));
}
@Override
public long valueForKey(long key) {
return timeZoneRounding.valueForKey(key);
}
@Override
@ -483,8 +513,13 @@ public abstract class TimeZoneRounding implements Rounding {
}
@Override
public long round(long utcMillis) {
return postOffset + timeZoneRounding.round(utcMillis + preOffset);
public long roundKey(long utcMillis) {
return timeZoneRounding.roundKey(utcMillis + preOffset);
}
@Override
public long valueForKey(long key) {
return postOffset + timeZoneRounding.valueForKey(key);
}
@Override

4
src/main/java/org/elasticsearch/search/aggregations/bucket/histogram/HistogramAggregator.java
View File

@ -87,7 +87,7 @@ public class HistogramAggregator extends BucketsAggregator {
long previousKey = Long.MIN_VALUE;
for (int i = 0; i < valuesCount; ++i) {
long value = values.nextValue();
long key = rounding.round(value);
long key = rounding.roundKey(value);
assert key >= previousKey;
if (key == previousKey) {
continue;
@ -110,7 +110,7 @@ public class HistogramAggregator extends BucketsAggregator {
if (ord < 0) {
continue; // slot is not allocated
}
buckets.add(histogramFactory.createBucket(bucketOrds.key(i), bucketDocCount(ord), bucketAggregations(ord)));
buckets.add(histogramFactory.createBucket(rounding.valueForKey(bucketOrds.key(i)), bucketDocCount(ord), bucketAggregations(ord)));
}