[
https://issues.apache.org/jira/browse/CASSANDRA-15215?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17436519#comment-17436519
]
Aleksandr Sorokoumov commented on CASSANDRA-15215:
--------------------------------------------------
[~benedict] if I understood your idea correctly, you suggest writing relevant
bytes directly instead of preparing the thread-local byte array and memcpy'ing
it into the given buffer. Is my interpretation correct? In the title and
description you also mentioned reads, but I haven't figured how to adjust this
idea there, so here are results for writes only.
h2. Code
||Branch||Description||
|[baseline|https://github.com/apache/cassandra/compare/trunk...Gerrrr:15215-baseline-trunk?expand=1]|trunk
+ benchmark|
|[patch|https://github.com/apache/cassandra/compare/trunk...Gerrrr:15215-trunk?expand=1]|patch
+ benchmark|
h2. Setup
Each benchmark does {{VIntCoding.writeUnsignedVInt}} on a {{long}} from 1 to 9
bytes. The write target is a {{ByteBuffer}} - both on- and off- heap.
The results are produced on a MBP 2019 - 2,3 GHz 8-Core Intel Core i9.
h2. Results
h3. Baseline
{noformat}
Benchmark (allocation) Mode Cnt Score Error Units
VIntCodingBench.testWrite1Byte HEAP avgt 15 9.084 ± 5.196 ns/op
VIntCodingBench.testWrite1Byte DIRECT avgt 15 5.037 ± 0.638 ns/op
VIntCodingBench.testWrite2Bytes HEAP avgt 15 15.604 ± 0.646 ns/op
VIntCodingBench.testWrite2Bytes DIRECT avgt 15 15.028 ± 0.568 ns/op
VIntCodingBench.testWrite3Bytes HEAP avgt 15 16.704 ± 0.461 ns/op
VIntCodingBench.testWrite3Bytes DIRECT avgt 15 17.410 ± 0.489 ns/op
VIntCodingBench.testWrite4Bytes HEAP avgt 15 17.086 ± 0.527 ns/op
VIntCodingBench.testWrite4Bytes DIRECT avgt 15 20.307 ± 0.705 ns/op
VIntCodingBench.testWrite5Bytes HEAP avgt 15 17.395 ± 0.578 ns/op
VIntCodingBench.testWrite5Bytes DIRECT avgt 15 17.558 ± 0.512 ns/op
VIntCodingBench.testWrite6Bytes HEAP avgt 15 18.114 ± 0.967 ns/op
VIntCodingBench.testWrite6Bytes DIRECT avgt 15 19.023 ± 0.591 ns/op
VIntCodingBench.testWrite7Bytes HEAP avgt 15 18.004 ± 0.298 ns/op
VIntCodingBench.testWrite7Bytes DIRECT avgt 15 19.081 ± 0.601 ns/op
VIntCodingBench.testWrite8Bytes HEAP avgt 15 18.466 ± 0.463 ns/op
VIntCodingBench.testWrite8Bytes DIRECT avgt 15 20.228 ± 5.620 ns/op
VIntCodingBench.testWrite9Bytes HEAP avgt 15 18.553 ± 0.537 ns/op
VIntCodingBench.testWrite9Bytes DIRECT avgt 15 20.101 ± 0.476 ns/op
{noformat}
h3. Patch
{noformat}
Benchmark (allocation) Mode Cnt Score Error Units
VIntCodingBench.testWrite1Byte HEAP avgt 15 4.728 ± 0.077 ns/op
VIntCodingBench.testWrite1Byte DIRECT avgt 15 6.415 ± 3.157 ns/op
VIntCodingBench.testWrite2Bytes HEAP avgt 15 8.244 ± 0.440 ns/op
VIntCodingBench.testWrite2Bytes DIRECT avgt 15 9.136 ± 3.979 ns/op
VIntCodingBench.testWrite3Bytes HEAP avgt 15 8.714 ± 0.134 ns/op
VIntCodingBench.testWrite3Bytes DIRECT avgt 15 9.690 ± 2.735 ns/op
VIntCodingBench.testWrite4Bytes HEAP avgt 15 8.634 ± 0.164 ns/op
VIntCodingBench.testWrite4Bytes DIRECT avgt 15 6.830 ± 0.061 ns/op
VIntCodingBench.testWrite5Bytes HEAP avgt 15 8.389 ± 0.207 ns/op
VIntCodingBench.testWrite5Bytes DIRECT avgt 15 8.059 ± 1.537 ns/op
VIntCodingBench.testWrite6Bytes HEAP avgt 15 10.861 ± 0.336 ns/op
VIntCodingBench.testWrite6Bytes DIRECT avgt 15 9.816 ± 1.482 ns/op
VIntCodingBench.testWrite7Bytes HEAP avgt 15 11.045 ± 0.419 ns/op
VIntCodingBench.testWrite7Bytes DIRECT avgt 15 10.702 ± 2.377 ns/op
VIntCodingBench.testWrite8Bytes HEAP avgt 15 10.375 ± 0.423 ns/op
VIntCodingBench.testWrite8Bytes DIRECT avgt 15 7.237 ± 0.176 ns/op
VIntCodingBench.testWrite9Bytes HEAP avgt 15 11.200 ± 0.365 ns/op
VIntCodingBench.testWrite9Bytes DIRECT avgt 15 8.152 ± 0.282 ns/op
{noformat}
> VIntCoding should read and write more efficiently
> -------------------------------------------------
>
> Key: CASSANDRA-15215
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15215
> Project: Cassandra
> Issue Type: Bug
> Components: Local/Compaction, Local/SSTable
> Reporter: Benedict Elliott Smith
> Priority: Normal
> Fix For: 3.0.x, 3.11.x, 4.x
>
>
> Most vints occupy significantly fewer than 8 bytes, and most buffers have >=
> 8 bytes spare, in which case we can construct the relevant bytes in a
> register and memcpy them to the correct position. Since we read and write a
> lot of vints, this waste is probably measurable, particularly during
> compaction and flush, and can probably be considered a performance bug.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
