mayya-sharipova

open_ai track

open_ai track modified: 1) exclude for merges to finish 2) include force-merge

Track progression:

• standalone indexing with 5 clients
• standalone search with 1 client
• standalone search with 8 clients
• parallel indexing with 1 client (target throughput of 500 docs/s)
• parallel search with 3 clients (target throughput of 100 op/s)
• force-merge

mayya-sharipova

open_ai track

Track progression:

• standalone indexing with 5 clients
• standalone search with 1 client
• standalone search with 8 clients
• parallel indexing with 1 client (target throughput of 500 docs/s)
• parallel search with 3 clients (target throughput of 100 op/s)

Baseline VS Candidate

• baseline: lucene_snapshot branch of Elasticsearch (with MADV_RANDOM feature flag)

mayya-sharipova

Metric	Task	Baseline	Contender	Diff	Unit	Diff %
Cumulative indexing time of primary shards		30.6003	31.6822	1.08185	min	+3.54%
Min cumulative indexing time across primary shard		15.2671	15.7348	0.46777	min	+3.06%
Median cumulative indexing time across primary shard		15.3002	15.8411	0.54092	min	+3.54%
Max cumulative indexing time across primary shard

mayya-sharipova

Metric	Task	Baseline	Contender	Diff	Unit	Diff %
Cumulative indexing time of primary shards		9.5825	9.45558	-0.12692	min	-1.32%
Min cumulative indexing time across primary shard		4.77432	4.70707	-0.06725	min	-1.41%
Median cumulative indexing time across primary shard		4.79125	4.72779	-0.06346	min	-1.32%
Max cumulative indexing time across primary shard

mayya-sharipova

Investigation on merge policy for vectors

Merge policy has been modified for vectors:

• Instead of merging segments of equal size, a merge is composed of the current largest segment with smallest segments
• Instead of choosing a merge with the smallest skew (preference was given to merges of segments of same size), preference is given to a merge of biggest segment with smallest segments.

These modifications were done to take advantage of optimization during merging that keeps the largest graph from merging segments and appends vectors from other merging segments to it.

Experiments

mayya-sharipova

Investigation on merge policy for vectors

Merge Policies:

• baseline: current TieredMergePolicy on Lucene main (floor_segment_size=2MB, maxMergeAtOnce=10, segsPerTier=10)
• candidate1: the same TieredMergePolicy with floor_segment_size=200MB
• candidate2: the same TieredMergePolicy floor_segment_size=200MB, maxMergeAtOnce=20, segsPerTier=20

Experiment 1

• 10M vectors of 100 dims
• NRT reader opens every second (to simulate concurrent with indexing searches)

mayya-sharipova

• 5Gb heap
• 2Gb for indexing buffer
• glove-100-angular M:16 efConstruction:100

recall dropped by 3-8%; QPS dropped by 30-50%

	9.3 recall	9.3 QPS	9.4 recall	9.4 QPS
n_cands=10	0.620	2745.933	0.565	1856.128

mayya-sharipova

import jdk.incubator.vector.FloatVector;
import jdk.incubator.vector.VectorSpecies;
import java.util.Random;

import static jdk.incubator.vector.VectorOperators.ADD;
import static jdk.incubator.vector.FloatVector.SPECIES_256;

public class TestVectorsAPI {
  private static final int SIZE = 128;
  private static float[] v1 = new float[SIZE];

mayya-sharipova

baseline

Profiler for cpu:
PROFILE SUMMARY from 921352 events (total: 921352)
  tests.profile.mode=cpu
  tests.profile.count=30
  tests.profile.stacksize=1
  tests.profile.linenumbers=false
PERCENT       CPU SAMPLES   STACK
12.85%        118371        java.io.BufferedOutputStream#write()

mayya-sharipova

baseline

Profiler for cpu:
PROFILE SUMMARY from 107744 events (total: 107744)
  tests.profile.mode=cpu
  tests.profile.count=30
  tests.profile.stacksize=1
  tests.profile.linenumbers=false
PERCENT       CPU SAMPLES   STACK