Skip to content

Commit

Permalink
Add CompressionLevel Calculation for PQ
Browse files Browse the repository at this point in the history
Currently, for product quantization, we set the calculated compression
level to NOT_CONFIGURED. The main issue with this is that if a user sets
up a disk-based index with PQ, no re-scoring will happen by default.

This change adds the calculation so that the proper re-scoring will
happen. The formula is fairly straightforward =>
actual compression = (d * 32) / (m * code_size). Then, we round to the
neareste compression level (because we only support discrete compression
levels).

One small issue with this is that if PQ is configured to have
compression > 32x, the value will be 32x. Functionally, the only issue
will be that we may not be as aggressive on oversampling for on disk
mode.

Signed-off-by: John Mazanec <[email protected]>
  • Loading branch information
jmazanec15 committed Oct 10, 2024
1 parent 5a56829 commit 907f1ec
Show file tree
Hide file tree
Showing 9 changed files with 181 additions and 62 deletions.
1 change: 1 addition & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
* KNNIterators should support with and without filters [#2155](https://github.com/opensearch-project/k-NN/pull/2155)
* Adding Support to Enable/Disble Share level Rescoring and Update Oversampling Factor[#2172](https://github.com/opensearch-project/k-NN/pull/2172)
* Add support to build vector data structures greedily and perform exact search when there are no engine files [#1942](https://github.com/opensearch-project/k-NN/issues/1942)
* Add CompressionLevel Calculation for PQ [#2200](https://github.com/opensearch-project/k-NN/pull/2200)
### Bug Fixes
* Add DocValuesProducers for releasing memory when close index [#1946](https://github.com/opensearch-project/k-NN/pull/1946)
* KNN80DocValues should only be considered for BinaryDocValues fields [#2147](https://github.com/opensearch-project/k-NN/pull/2147)
Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,92 @@
/*
* Copyright OpenSearch Contributors
* SPDX-License-Identifier: Apache-2.0
*/

package org.opensearch.knn.index.engine.faiss;

import org.opensearch.common.ValidationException;
import org.opensearch.knn.index.engine.Encoder;
import org.opensearch.knn.index.engine.KNNMethodConfigContext;
import org.opensearch.knn.index.engine.MethodComponentContext;
import org.opensearch.knn.index.mapper.CompressionLevel;

import static org.opensearch.knn.common.KNNConstants.ENCODER_PARAMETER_PQ_CODE_SIZE;
import static org.opensearch.knn.common.KNNConstants.ENCODER_PARAMETER_PQ_M;

/**
* Abstract class for Faiss PQ encoders. This class provides the common logic for product quantization based encoders
*/
public abstract class AbstractFaissPQEncoder implements Encoder {

@Override
public CompressionLevel calculateCompressionLevel(
MethodComponentContext methodComponentContext,
KNNMethodConfigContext knnMethodConfigContext
) {
// Roughly speaking, PQ can be configured to produce a lot of different compression levels. The "m" parameter
// specifies how many sub-vectors to break the vector up into, and then the "code_size" represents the number
// of bits to encode each subvector. Thus, a d-dimensional vector of float32s goes from
// d*32 -> (m)*code_size bits. So if we want (d*32)/(m*code_size) will be the compression level.
//
// Example:
// d=768, m=384, code_size=8
// (768*32)/(384*8) = 8x (i.e. 24,576 vs. 3,072).
//
// Because of this variability, we will need to properly round to one of the supported values.
if (methodComponentContext.getParameters().containsKey(ENCODER_PARAMETER_PQ_M) == false
|| methodComponentContext.getParameters().containsKey(ENCODER_PARAMETER_PQ_CODE_SIZE) == false) {
return CompressionLevel.NOT_CONFIGURED;
}

// Map the number of bits passed in, back to the compression level
Object value = methodComponentContext.getParameters().get(ENCODER_PARAMETER_PQ_M);
ValidationException validationException = getMethodComponent().getParameters()
.get(ENCODER_PARAMETER_PQ_M)
.validate(value, knnMethodConfigContext);
if (validationException != null) {
throw validationException;
}
Integer m = (Integer) value;
value = methodComponentContext.getParameters().get(ENCODER_PARAMETER_PQ_CODE_SIZE);
validationException = getMethodComponent().getParameters()
.get(ENCODER_PARAMETER_PQ_CODE_SIZE)
.validate(value, knnMethodConfigContext);
if (validationException != null) {
throw validationException;
}
Integer codeSize = (Integer) value;
int dimension = knnMethodConfigContext.getDimension();

float actualCompression = ((float) dimension * 32) / (m * codeSize);

if (actualCompression < 2.0f) {
return CompressionLevel.x1;
}

if (actualCompression < 4.0f) {
return CompressionLevel.x2;
}

if (actualCompression < 8.0f) {
return CompressionLevel.x4;
}

if (actualCompression < 16.0f) {
return CompressionLevel.x8;
}

if (actualCompression < 32.0f) {
return CompressionLevel.x16;
}

if (actualCompression < 64.0f) {
return CompressionLevel.x32;
}

// TODO: The problem is that the theoretical compression level of PQ can be in the thousands. Thus, Im not sure
// it makes sense to have an enum all the way up to that value. So, for now, we will just return the max
// compression
return CompressionLevel.MAX_COMPRESSION_LEVEL;
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -8,12 +8,8 @@
import com.google.common.collect.ImmutableSet;
import org.opensearch.knn.common.KNNConstants;
import org.opensearch.knn.index.VectorDataType;
import org.opensearch.knn.index.engine.Encoder;
import org.opensearch.knn.index.engine.KNNMethodConfigContext;
import org.opensearch.knn.index.engine.MethodComponent;
import org.opensearch.knn.index.engine.MethodComponentContext;
import org.opensearch.knn.index.engine.Parameter;
import org.opensearch.knn.index.mapper.CompressionLevel;

import java.util.Objects;
import java.util.Set;
Expand All @@ -30,7 +26,7 @@
* Faiss HNSW PQ encoder. Right now, the implementations are slightly different during validation between this an
* {@link FaissIVFPQEncoder}. Hence, they are separate classes.
*/
public class FaissHNSWPQEncoder implements Encoder {
public class FaissHNSWPQEncoder extends AbstractFaissPQEncoder {

private static final Set<VectorDataType> SUPPORTED_DATA_TYPES = ImmutableSet.of(VectorDataType.FLOAT);

Expand Down Expand Up @@ -72,13 +68,4 @@ public class FaissHNSWPQEncoder implements Encoder {
public MethodComponent getMethodComponent() {
return METHOD_COMPONENT;
}

@Override
public CompressionLevel calculateCompressionLevel(
MethodComponentContext methodComponentContext,
KNNMethodConfigContext knnMethodConfigContext
) {
// TODO: For now, not supported out of the box
return CompressionLevel.NOT_CONFIGURED;
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -5,15 +5,12 @@

package org.opensearch.knn.index.engine.faiss;

import com.google.common.annotations.VisibleForTesting;
import com.google.common.collect.ImmutableSet;
import org.opensearch.knn.common.KNNConstants;
import org.opensearch.knn.index.VectorDataType;
import org.opensearch.knn.index.engine.Encoder;
import org.opensearch.knn.index.engine.KNNMethodConfigContext;
import org.opensearch.knn.index.engine.MethodComponent;
import org.opensearch.knn.index.engine.MethodComponentContext;
import org.opensearch.knn.index.engine.Parameter;
import org.opensearch.knn.index.mapper.CompressionLevel;

import java.util.Set;

Expand All @@ -30,11 +27,12 @@
* Faiss IVF PQ encoder. Right now, the implementations are slightly different during validation between this an
* {@link FaissHNSWPQEncoder}. Hence, they are separate classes.
*/
public class FaissIVFPQEncoder implements Encoder {
public class FaissIVFPQEncoder extends AbstractFaissPQEncoder {

private static final Set<VectorDataType> SUPPORTED_DATA_TYPES = ImmutableSet.of(VectorDataType.FLOAT);

private final static MethodComponent METHOD_COMPONENT = MethodComponent.Builder.builder(KNNConstants.ENCODER_PQ)
@VisibleForTesting
final static MethodComponent METHOD_COMPONENT = MethodComponent.Builder.builder(KNNConstants.ENCODER_PQ)
.addSupportedDataTypes(SUPPORTED_DATA_TYPES)
.addParameter(
ENCODER_PARAMETER_PQ_M,
Expand Down Expand Up @@ -93,13 +91,4 @@ public class FaissIVFPQEncoder implements Encoder {
public MethodComponent getMethodComponent() {
return METHOD_COMPONENT;
}

@Override
public CompressionLevel calculateCompressionLevel(
MethodComponentContext methodComponentContext,
KNNMethodConfigContext knnMethodConfigContext
) {
// TODO: For now, not supported out of the box
return CompressionLevel.NOT_CONFIGURED;
}
}
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,8 @@ public enum CompressionLevel {
x16(16, "16x", new RescoreContext(3.0f, false), Set.of(Mode.ON_DISK)),
x32(32, "32x", new RescoreContext(3.0f, false), Set.of(Mode.ON_DISK));

public static final CompressionLevel MAX_COMPRESSION_LEVEL = CompressionLevel.x32;

// Internally, an empty string is easier to deal with them null. However, from the mapping,
// we do not want users to pass in the empty string and instead want null. So we make the conversion here
public static final String[] NAMES_ARRAY = new String[] {
Expand Down
3 changes: 2 additions & 1 deletion src/test/java/org/opensearch/knn/index/FaissIT.java
Original file line number Diff line number Diff line change
Expand Up @@ -1714,7 +1714,8 @@ public void testIVF_InvalidPQM_thenFail() {
() -> ingestDataAndTrainModel(modelId, trainingIndexName, trainingFieldName, dimension, modelDescription, in, trainingDataCount)
);
assertTrue(
re.getMessage().contains("Validation Failed: 1: parameter validation failed for MethodComponentContext parameter [encoder].;")
re.getMessage(),
re.getMessage().contains("Validation Failed: 1: parameter validation failed for Integer parameter [m].;")
);
}

Expand Down
Original file line number Diff line number Diff line change
@@ -0,0 +1,79 @@
/*
* Copyright OpenSearch Contributors
* SPDX-License-Identifier: Apache-2.0
*/

package org.opensearch.knn.index.engine.faiss;

import lombok.SneakyThrows;
import org.opensearch.knn.KNNTestCase;
import org.opensearch.knn.index.engine.Encoder;
import org.opensearch.knn.index.engine.KNNMethodConfigContext;
import org.opensearch.knn.index.engine.MethodComponent;
import org.opensearch.knn.index.engine.MethodComponentContext;
import org.opensearch.knn.index.mapper.CompressionLevel;

import java.util.Map;

import static org.opensearch.knn.common.KNNConstants.ENCODER_PARAMETER_PQ_CODE_SIZE;
import static org.opensearch.knn.common.KNNConstants.ENCODER_PARAMETER_PQ_M;
import static org.opensearch.knn.common.KNNConstants.ENCODER_PQ;

public class AbstractFaissPQEncoderTests extends KNNTestCase {

@SneakyThrows
public void testCalculateCompressionLevel() {
AbstractFaissPQEncoder encoder = new AbstractFaissPQEncoder() {
@Override
public MethodComponent getMethodComponent() {
return FaissIVFPQEncoder.METHOD_COMPONENT;
}
};

// Compression formula is:
// actual_compression = (d*32)/(m*code_size) and then round down to nearest: 1x, 2x, 4x, 8x, 16x, 32x

// d=768
// m=2
// code_size=8
// actual_compression = (768*32)/(2*8) = 1,536x
// expected_compression = Max compression level
assertCompressionLevel(2, 8, 768, CompressionLevel.MAX_COMPRESSION_LEVEL, encoder);

// d=32
// m=4
// code_size=16
// actual_compression = (32*32)/(4*16) = 16x
// expected_compression = Max compression level
assertCompressionLevel(4, 16, 32, CompressionLevel.x16, encoder);

// d=1536
// m=768
// code_size=8
// actual_compression = (1536*32)/(768*8) = 8x
// expected_compression = Max compression level
assertCompressionLevel(768, 8, 1536, CompressionLevel.x8, encoder);

// d=128
// m=128
// code_size=8
// actual_compression = (128*32)/(128*8) = 4x
// expected_compression = Max compression level
assertCompressionLevel(128, 8, 128, CompressionLevel.x4, encoder);
}

private void assertCompressionLevel(int m, int codeSize, int d, CompressionLevel expectedCompression, Encoder encoder) {
assertEquals(
expectedCompression,
encoder.calculateCompressionLevel(generateMethodComponentContext(m, codeSize), generateKNNMethodConfigContext(d))
);
}

private MethodComponentContext generateMethodComponentContext(int m, int codeSize) {
return new MethodComponentContext(ENCODER_PQ, Map.of(ENCODER_PARAMETER_PQ_M, m, ENCODER_PARAMETER_PQ_CODE_SIZE, codeSize));
}

private KNNMethodConfigContext generateKNNMethodConfigContext(int dimension) {
return KNNMethodConfigContext.builder().dimension(dimension).build();
}
}

This file was deleted.

This file was deleted.

0 comments on commit 907f1ec

Please sign in to comment.