Optimize BytesArray::indexOf, which is used heavily in ndjson parsing #135087
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ChrisHegarty
added
:Core/Infra/Core
|
Pinging @elastic/es-perf (Team:Performance) |
|
Pinging @elastic/es-core-infra (Team:Core/Infra) |
|
Hi @ChrisHegarty, I've created a changelog YAML for you. |
ChrisHegarty
commented
Sep 19, 2025
libs/simdvec/src/main/java/org/elasticsearch/simdvec/internal/vectorization/ByteArrayUtils.java
Show resolved
Hide resolved
benwtrent
reviewed
Sep 19, 2025
benchmarks/src/main/java/org/elasticsearch/benchmark/bytes/BytesArrayIndexOfBenchmark.java
Show resolved
Hide resolved
.../main21/java/org/elasticsearch/simdvec/internal/vectorization/PanamaESVectorUtilSupport.java
Show resolved
Hide resolved
benchmarks/src/main/java/org/elasticsearch/benchmark/bytes/BytesArrayIndexOfBenchmark.java
Outdated
Show resolved
Hide resolved
|
Seems like, maybe we don't need to check the length at all? This is surprising to me. But for index size of 64 (way too small to be Running on my macbook (128 arm) and its looking nice! |
ha! Thanks for checking this. Lemme add those sizes to the benchmark and do some more runs. But otherwise looks like we don't need to fall back to the old code - as it's slower. |
|
Benchmark summary: Panama consistently outperforms the baseline on all architectures. Benchmark results: Linux 6.8.0-1029-aws aarch64 AWS Graviton3 Linux 6.14.0-1010-aws x86_64 Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz Linux 6.8.0-79-generic x86_64 11th Gen Intel(R) Core(TM) i5-11400 @ 2.60GHz |
benwtrent
approved these changes
Sep 19, 2025
.../main21/java/org/elasticsearch/simdvec/internal/vectorization/PanamaESVectorUtilSupport.java
Outdated
Show resolved
Hide resolved
|
Latest results; summary as good or better than the previous set. Small sizes with tails have improved. Linux 6.8.0-1029-aws aarch64 AWS Graviton3 Linux 6.14.0-1010-aws x86_64 Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz Linux 6.8.0-79-generic x86_64 11th Gen Intel(R) Core(TM) i5-11400 @ 2.60GHz |
ldematte
reviewed
Sep 22, 2025
| } | ||
|
|
||
| @Override | ||
| public int indexOf(final byte[] bytes, final int offset, final int length, final byte marker) { |
There was a problem hiding this comment.
Nit: I don't think we need to keep this in main21, as 21 is the minimum version we support now. We need to apply the mrjar plugin for the preview bit handling, but no need for a different sourceset I think
gmjehovich
pushed a commit
to gmjehovich/elasticsearch
that referenced
this pull request
Sep 22, 2025
…elastic#135087) This commit optimizes BytesArray::indexOf, which is used heavily in ndjson parsing. I simply extracted the indexOf functionality into ESVectorUtil, so that we could have a default implementation (the current one), and a Panama Vectorized one. BytesArray::indexOf is used heavily for newline json delimited parsing, which backs the _bulk API. Here’s the result of the benchmark run on Linux x64, Intel SkyLake: Size Baseline ops/ms Panama ops/ms Speedup 4,096 2,439 41,689 ~17.1x 16,384 624 12,297 ~19.7x 65,536 156 1,689 ~10.8x 1,048,576 9.8 73.4 ~7.5x The Panama version is 7–20× faster than the baseline depending on input size. For small and medium arrays (4 KB and 16 KB), the speedup is dramatic (~17–20×), showing that the vectorization pays off most when the data set fits easily in cache. For larger arrays (64 KB – 1 MB), the relative advantage drops (~10× at 64 KB, ~7.5× at 1 MB). That’s expected - as array size grows, the bottleneck shifts from CPU to memory bandwidth, reducing the benefit of SIMD. In short: Panama delivers an order-of-magnitude improvement over the scalar implementation, especially for small-to-medium buffers, while still offering a solid multiple (~7×) speedup even on very large buffers. The Panama implementation is between 10-6x faster on AVX2, and ~5x times faster on ARM.
DonalEvans
pushed a commit
to DonalEvans/elasticsearch
that referenced
this pull request
Sep 22, 2025
…elastic#135087) This commit optimizes BytesArray::indexOf, which is used heavily in ndjson parsing. I simply extracted the indexOf functionality into ESVectorUtil, so that we could have a default implementation (the current one), and a Panama Vectorized one. BytesArray::indexOf is used heavily for newline json delimited parsing, which backs the _bulk API. Here’s the result of the benchmark run on Linux x64, Intel SkyLake: Size Baseline ops/ms Panama ops/ms Speedup 4,096 2,439 41,689 ~17.1x 16,384 624 12,297 ~19.7x 65,536 156 1,689 ~10.8x 1,048,576 9.8 73.4 ~7.5x The Panama version is 7–20× faster than the baseline depending on input size. For small and medium arrays (4 KB and 16 KB), the speedup is dramatic (~17–20×), showing that the vectorization pays off most when the data set fits easily in cache. For larger arrays (64 KB – 1 MB), the relative advantage drops (~10× at 64 KB, ~7.5× at 1 MB). That’s expected - as array size grows, the bottleneck shifts from CPU to memory bandwidth, reducing the benefit of SIMD. In short: Panama delivers an order-of-magnitude improvement over the scalar implementation, especially for small-to-medium buffers, while still offering a solid multiple (~7×) speedup even on very large buffers. The Panama implementation is between 10-6x faster on AVX2, and ~5x times faster on ARM.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Optimize BytesArray::indexOf, which is used heavily in ndjson parsing.
I simply extracted the indexOf functionality into ESVectorUtil, so that we could have a default implementation (the current one), and a Panama Vectorized one. BytesArray::indexOf is used heavily for newline json delimited parsing, which backs the _bulk API.
Here’s the result of the benchmark run on Linux x64, Intel SkyLake:
The Panama version is 7–20× faster than the baseline depending on input size.
For small and medium arrays (4 KB and 16 KB), the speedup is dramatic (~17–20×), showing that the vectorization pays off most when the data set fits easily in cache.
For larger arrays (64 KB – 1 MB), the relative advantage drops (~10× at 64 KB, ~7.5× at 1 MB). That’s expected - as array size grows, the bottleneck shifts from CPU to memory bandwidth, reducing the benefit of SIMD.
In short: Panama delivers an order-of-magnitude improvement over the scalar implementation, especially for small-to-medium buffers, while still offering a solid multiple (~7×) speedup even on very large buffers.
The Panama implementation is between 10-6x faster on AVX2, and ~5x times faster on ARM.