Revert "LUCENE-8374 part 3/4: Reduce reads for sparse DocValues".

LUCENE-8374 was committed without consensus and is expected to be superseded by LUCENE-8585.

This reverts commit 7949b98f80.
This commit is contained in:
Toke Eskildsen 2018-12-11 14:14:07 +01:00
parent 3158d0c485
commit 6c5d87a505
6 changed files with 114 additions and 232 deletions

View File

@ -303,9 +303,6 @@ final class IndexedDISI extends DocIdSetIterator {
final int targetInBlock = target & 0xFFFF;
final int targetWordIndex = targetInBlock >>> 6;
// If possible, skip ahead using the rank cache
disi.rankSkip(disi, target);
for (int i = disi.wordIndex + 1; i <= targetWordIndex; ++i) {
disi.word = disi.slice.readLong();
disi.numberOfOnes += Long.bitCount(disi.word);
@ -340,9 +337,6 @@ final class IndexedDISI extends DocIdSetIterator {
final int targetInBlock = target & 0xFFFF;
final int targetWordIndex = targetInBlock >>> 6;
// If possible, skip ahead using the rank cache
disi.rankSkip(disi, target);
for (int i = disi.wordIndex + 1; i <= targetWordIndex; ++i) {
disi.word = disi.slice.readLong();
disi.numberOfOnes += Long.bitCount(disi.word);
@ -379,47 +373,4 @@ final class IndexedDISI extends DocIdSetIterator {
abstract boolean advanceExactWithinBlock(IndexedDISI disi, int target) throws IOException;
}
/**
* If the distance between the current position and the target is > 8 words, the rank cache will
* be used to guarantee a worst-case of 1 rank-lookup and 7 word-read-and-count-bits operations.
* Note: This does not guarantee a skip up to target, only up to nearest rank boundary. It is the
* responsibility of the caller to iterate further to reach target.
* @param disi standard DISI.
* @param target the wanted docID for which to calculate set-flag and index.
* @throws IOException if a disi seek failed.
*/
private void rankSkip(IndexedDISI disi, int target) throws IOException {
final int targetInBlock = target & 0xFFFF;
final int targetWordIndex = targetInBlock >>> 6;
// If the distance between the current position and the target is >= 8
// then it pays to use the rank to jump
if (!(disi.cache.hasRank() && targetWordIndex - disi.wordIndex >= IndexedDISICache.RANK_BLOCK_LONGS)) {
return;
}
int rankPos = disi.cache.denseRankPosition(target);
if (rankPos == -1) {
return;
}
int rank = disi.cache.getRankInBlock(rankPos);
if (rank == -1) {
return;
}
int rankIndex = disi.denseOrigoIndex + rank;
int rankWordIndex = (rankPos & 0xFFFF) >> 6;
long rankOffset = disi.blockStart + 4 + (rankWordIndex * 8);
long mark = disi.slice.getFilePointer();
disi.slice.seek(rankOffset);
long rankWord = disi.slice.readLong();
int rankNOO = rankIndex + Long.bitCount(rankWord);
rankOffset += Long.BYTES;
//disi.slice.seek(mark);
disi.wordIndex = rankWordIndex;
disi.word = rankWord;
disi.numberOfOnes = rankNOO;
}
}

View File

@ -44,8 +44,6 @@ public class IndexedDISICacheFactory implements Accountable {
// jump-table and rank for DISI blocks
private final Map<Long, IndexedDISICache> disiPool = new HashMap<>();
// jump-table for numerics with variable bits per value (dates, longs...)
private final Map<String, VaryingBPVJumpTable> vBPVPool = new HashMap<>();
/**
* Create a cached {@link IndexedDISI} instance.
@ -76,24 +74,6 @@ public class IndexedDISICacheFactory implements Accountable {
return new IndexedDISI(data, offset, length, cost, cache, name);
}
/**
* Creates a cache (jump table) for variable bits per value numerics and returns it.
* If the cache has previously been created, the old cache is returned.
* @param name the name for the cache, typically the field name. Used as key for later retrieval.
* @param slice the long values with varying bits per value.
* @param valuesLength the length in bytes of the slice.
* @return a jump table for the longs in the given slice or null if the structure is not suitable for caching.
*/
VaryingBPVJumpTable getVBPVJumpTable(String name, RandomAccessInput slice, long valuesLength) throws IOException {
VaryingBPVJumpTable jumpTable = vBPVPool.get(name);
if (jumpTable == null) {
// TODO: Avoid overlapping builds of the same jump table for performance reasons
jumpTable = new VaryingBPVJumpTable(slice, name, valuesLength);
vBPVPool.put(name, jumpTable);
}
return jumpTable;
}
/**
* Creates a cache (jump table) for {@link IndexedDISI}.
* If the cache has previously been created, the old cache is returned.
@ -153,26 +133,16 @@ public class IndexedDISICacheFactory implements Accountable {
public long getDISIBlocksWithRankCount() {
return disiPool.values().stream().filter(IndexedDISICache::hasRank).count();
}
public long getVaryingBPVCount() {
return vBPVPool.size();
}
@Override
public long ramBytesUsed() {
long mem = RamUsageEstimator.shallowSizeOf(this) +
RamUsageEstimator.shallowSizeOf(disiPool) +
RamUsageEstimator.shallowSizeOf(vBPVPool);
RamUsageEstimator.shallowSizeOf(disiPool);
for (Map.Entry<Long, IndexedDISICache> cacheEntry: disiPool.entrySet()) {
mem += RamUsageEstimator.shallowSizeOf(cacheEntry);
mem += RamUsageEstimator.sizeOf(cacheEntry.getKey());
mem += cacheEntry.getValue().ramBytesUsed();
}
for (Map.Entry<String, VaryingBPVJumpTable> cacheEntry: vBPVPool.entrySet()) {
String key = cacheEntry.getKey();
mem += RamUsageEstimator.shallowSizeOf(cacheEntry);
mem += RamUsageEstimator.shallowSizeOf(key)+key.length()*2;
mem += cacheEntry.getValue().ramBytesUsed();
}
return mem;
}
@ -181,63 +151,5 @@ public class IndexedDISICacheFactory implements Accountable {
*/
void releaseAll() {
disiPool.clear();
vBPVPool.clear();
}
/**
* Jump table used by Lucene70DocValuesProducer.VaryingBPVReader to avoid iterating all blocks from
* current to wanted index for numerics with variable bits per value. The jump table holds offsets for all blocks.
*/
public static class VaryingBPVJumpTable implements Accountable {
// TODO: It is way overkill to use longs here for practically all indexes. Maybe a PackedInts representation?
long[] offsets = new long[10];
final String creationStats;
VaryingBPVJumpTable(RandomAccessInput slice, String name, long valuesLength) throws IOException {
final long startTime = System.nanoTime();
int block = -1;
long offset;
long blockEndOffset = 0;
int bitsPerValue;
do {
offset = blockEndOffset;
offsets = ArrayUtil.grow(offsets, block+2); // No-op if large enough
offsets[block+1] = offset;
bitsPerValue = slice.readByte(offset++);
offset += Long.BYTES; // Skip over delta as we do not resolve the values themselves at this point
if (bitsPerValue == 0) {
blockEndOffset = offset;
} else {
final int length = slice.readInt(offset);
offset += Integer.BYTES;
blockEndOffset = offset + length;
}
block++;
} while (blockEndOffset < valuesLength-Byte.BYTES);
offsets = ArrayUtil.copyOfSubArray(offsets, 0, block+1);
creationStats = String.format(Locale.ENGLISH,
"name=%s, blocks=%d, time=%dms",
name, offsets.length, (System.nanoTime()-startTime)/1000000);
}
/**
* @param block the logical block in the vBPV structure ( valueindex/16384 ).
* @return the index slice offset for the vBPV block (1 block = 16384 values) or -1 if not available.
*/
long getBlockOffset(long block) {
// Technically a limitation in caching vs. VaryingBPVReader to limit to 2b blocks of 16K values
return offsets[(int) block];
}
@Override
public long ramBytesUsed() {
return RamUsageEstimator.shallowSizeOf(this) +
(offsets == null ? 0 : RamUsageEstimator.sizeOf(offsets)) + // offsets
RamUsageEstimator.NUM_BYTES_OBJECT_HEADER + creationStats.length()*2; // creationStats
}
}
}

View File

@ -465,18 +465,44 @@ final class Lucene70DocValuesProducer extends DocValuesProducer implements Close
}
};
} else {
final RandomAccessInput slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
if (entry.blockShift >= 0) {
// dense but split into blocks of different bits per value
final int shift = entry.blockShift;
final long mul = entry.gcd;
final int mask = (1 << shift) - 1;
return new DenseNumericDocValues(maxDoc) {
final VaryingBPVReader vBPVReader = new VaryingBPVReader(entry);
int block = -1;
long delta;
long offset;
long blockEndOffset;
LongValues values;
@Override
public long longValue() throws IOException {
return vBPVReader.getLongValue(doc);
final int block = doc >>> shift;
if (this.block != block) {
int bitsPerValue;
do {
offset = blockEndOffset;
bitsPerValue = slice.readByte(offset++);
delta = slice.readLong(offset);
offset += Long.BYTES;
if (bitsPerValue == 0) {
blockEndOffset = offset;
} else {
final int length = slice.readInt(offset);
offset += Integer.BYTES;
blockEndOffset = offset + length;
}
this.block ++;
} while (this.block != block);
values = bitsPerValue == 0 ? LongValues.ZEROES : DirectReader.getInstance(slice, bitsPerValue, offset);
}
return mul * values.get(doc & mask) + delta;
}
};
} else {
final RandomAccessInput slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
final LongValues values = DirectReader.getInstance(slice, entry.bitsPerValue);
if (entry.table != null) {
final long[] table = entry.table;
@ -510,19 +536,45 @@ final class Lucene70DocValuesProducer extends DocValuesProducer implements Close
}
};
} else {
final RandomAccessInput slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
if (entry.blockShift >= 0) {
// sparse and split into blocks of different bits per value
final int shift = entry.blockShift;
final long mul = entry.gcd;
final int mask = (1 << shift) - 1;
return new SparseNumericDocValues(disi) {
final VaryingBPVReader vBPVReader = new VaryingBPVReader(entry);
int block = -1;
long delta;
long offset;
long blockEndOffset;
LongValues values;
@Override
public long longValue() throws IOException {
final int index = disi.index();
return vBPVReader.getLongValue(index);
final int block = index >>> shift;
if (this.block != block) {
int bitsPerValue;
do {
offset = blockEndOffset;
bitsPerValue = slice.readByte(offset++);
delta = slice.readLong(offset);
offset += Long.BYTES;
if (bitsPerValue == 0) {
blockEndOffset = offset;
} else {
final int length = slice.readInt(offset);
offset += Integer.BYTES;
blockEndOffset = offset + length;
}
this.block ++;
} while (this.block != block);
values = bitsPerValue == 0 ? LongValues.ZEROES : DirectReader.getInstance(slice, bitsPerValue, offset);
}
return mul * values.get(index & mask) + delta;
}
};
} else {
final RandomAccessInput slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
final LongValues values = DirectReader.getInstance(slice, entry.bitsPerValue);
if (entry.table != null) {
final long[] table = entry.table;
@ -556,20 +608,47 @@ final class Lucene70DocValuesProducer extends DocValuesProducer implements Close
}
};
} else {
final RandomAccessInput slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
if (entry.blockShift >= 0) {
final int shift = entry.blockShift;
final long mul = entry.gcd;
final long mask = (1L << shift) - 1;
return new LongValues() {
final VaryingBPVReader vBPVReader = new VaryingBPVReader(entry);
@Override
long block = -1;
long delta;
long offset;
long blockEndOffset;
LongValues values;
public long get(long index) {
final long block = index >>> shift;
if (this.block != block) {
assert block > this.block : "Reading backwards is illegal: " + this.block + " < " + block;
int bitsPerValue;
do {
offset = blockEndOffset;
try {
return vBPVReader.getLongValue(index);
bitsPerValue = slice.readByte(offset++);
delta = slice.readLong(offset);
offset += Long.BYTES;
if (bitsPerValue == 0) {
blockEndOffset = offset;
} else {
final int length = slice.readInt(offset);
offset += Integer.BYTES;
blockEndOffset = offset + length;
}
} catch (IOException e) {
throw new RuntimeException(e);
}
this.block ++;
} while (this.block != block);
values = bitsPerValue == 0 ? LongValues.ZEROES : DirectReader.getInstance(slice, bitsPerValue, offset);
}
return mul * values.get(index & mask) + delta;
}
};
} else {
final RandomAccessInput slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
final LongValues values = DirectReader.getInstance(slice, entry.bitsPerValue);
if (entry.table != null) {
final long[] table = entry.table;
@ -1378,64 +1457,4 @@ final class Lucene70DocValuesProducer extends DocValuesProducer implements Close
CodecUtil.checksumEntireFile(data);
}
/**
* Reader for longs split into blocks of different bits per values.
* The longs are requested by index and must be accessed in monotonically increasing order.
*/
// Note: The order requirement could be removed as the jump-tables allow for backwards iteration.
private class VaryingBPVReader {
final RandomAccessInput slice;
final NumericEntry entry;
final int shift;
final long mul;
final int mask;
long block = -1;
long delta;
long offset;
long blockEndOffset;
LongValues values;
VaryingBPVReader(NumericEntry entry) throws IOException {
this.entry = entry;
slice = data.randomAccessSlice(entry.valuesOffset, entry.valuesLength);
shift = entry.blockShift;
mul = entry.gcd;
mask = (1 << shift) - 1;
}
long getLongValue(long index) throws IOException {
final long block = index >>> shift;
if (this.block != block) {
int bitsPerValue;
do {
// If the needed block is the one directly following the current block, it is cheaper to avoid the cache
if (block != this.block+1) {
IndexedDISICacheFactory.VaryingBPVJumpTable cache;
if ((cache = disiCacheFactory.getVBPVJumpTable(entry.name, slice, entry.valuesLength)) != null) {
long candidateOffset;
if ((candidateOffset = cache.getBlockOffset(block)) != -1) {
blockEndOffset = candidateOffset;
this.block = block - 1;
}
}
}
offset = blockEndOffset;
bitsPerValue = slice.readByte(offset++);
delta = slice.readLong(offset);
offset += Long.BYTES;
if (bitsPerValue == 0) {
blockEndOffset = offset;
} else {
final int length = slice.readInt(offset);
offset += Integer.BYTES;
blockEndOffset = offset + length;
}
this.block++;
} while (this.block != block);
values = bitsPerValue == 0 ? LongValues.ZEROES : DirectReader.getInstance(slice, bitsPerValue, offset);
}
return mul * values.get(index & mask) + delta;
}
}
}