GH-3530: Optimize BYTE_STREAM_SPLIT encoding/decoding#3569
Open
iemejia wants to merge 1 commit into
Open
Conversation
This was referenced May 17, 2026
Reader: replace generic ByteBuffer.get() transpose loop in decodeData() with specialized single-pass loops for element sizes 2/4/8/12/16 bytes plus a stream-oriented generic fallback. Bulk-access the backing array directly when available, falling back to a single bulk copy for direct buffers. Writer: replace per-value scatterBytes() (which allocates a temp byte[] and issues N single-byte stream writes) with batched scatter buffers. Int/Long values accumulate in int[]/long[] batches of 64 and flush as bulk write(byte[], off, len) calls -- one per stream. FLBA uses per-stream byte[][] scratch buffers with the same batching strategy. getBufferedSize() now accounts for unflushed batch values. Add JMH benchmarks for scalar encode/decode of all 5 BSS types (FLOAT, DOUBLE, INT32, INT64, FIXED_LEN_BYTE_ARRAY). Add TestDataFactory for deterministic FLBA benchmark data generation. Add unit tests for transpose specializations, batch-boundary crossing, getBufferedSize with partial batches, direct ByteBuffer decode paths, and close/reset with pending unflushed batches.
iemejia
force-pushed
the
parquet-perf-v2-par5-bss
branch
from
f7fdee5 to
84de9c6
Compare
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Part of #3530 — Apache Parquet Java Performance Improvements
Summary
Optimize scalar encode/decode for the BYTE_STREAM_SPLIT encoding.
Reader: Specialized transpose loops for element sizes 2/4/8/12/16 bytes plus generic fallback. Bulk array access when backing array is available.
Writer: Batched scatter buffers (
int[]/long[]batches of 64) replacing per-valuescatterBytes()which allocated tempbyte[]and issued N single-byte writes.Includes unit tests for transpose specializations, batch-boundary crossing,
getBufferedSizewith partial batches, direct ByteBuffer decode paths, and close/reset with pending unflushed batches.JMH benchmarks:
BssEncodingBenchmark,BssDecodingBenchmarkcovering FLOAT, DOUBLE, INT32, INT64, and FIXED_LEN_BYTE_ARRAY.Benchmark results
Environment: JDK 25.0.3 (Temurin), OpenJDK 64-Bit Server VM, JMH 1.37, Linux x86_64.
Decoding:
Encoding:
Every benchmark shows clear improvement with no regressions. 8-byte types benefit most from the batched scatter (6.9x) since the baseline scattered 8 bytes per value into 8 separate streams.