[
https://issues.apache.org/jira/browse/CASSANDRA-16360?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17751773#comment-17751773
]
Maxim Muzafarov commented on CASSANDRA-16360:
---------------------------------------------
I have run the most promising options for us according to the state we now have
in the trunk branch (the Guava - hasherCrc32c and Snappy - pureJavaCrc32c are
not included in these results as it seems we can't benefit from them). So, the
_CRC32C_ works slightly better with JDK11 and significantly better with JDK17,
both of which ship with the _java.util.zip.CRC32C_ class. I also have run
benchmarks on the Apple M1 and got no difference, as expected.
h4. Configuration
h5. Java
{code}
openjdk 17.0.8 2023-07-18
OpenJDK Runtime Environment Temurin-17.0.8+7 (build 17.0.8+7)
{code}
{code}
openjdk 11.0.20 2023-07-18
OpenJDK Runtime Environment Temurin-11.0.20+8 (build 11.0.20+8)
OpenJDK 64-Bit Server VM Temurin-11.0.20+8 (build 11.0.20+8, mixed mode)
{code}
h5. CPU
{code}
model : 142
model name : Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz
cache size : 8192 KB
{code}
h4. Results
{code:title=JDK11}
Benchmark (bufferSize) Mode Cnt Score
Error Units
ChecksumBench.benchCrc32 31 avgt 5 70.069 ±
32.848 ns/op
ChecksumBench.benchCrc32 131 avgt 5 99.017 ±
8.763 ns/op
ChecksumBench.benchCrc32 517 avgt 5 80.374 ±
15.498 ns/op
ChecksumBench.benchCrc32 2041 avgt 5 170.778 ±
3.089 ns/op
ChecksumBench.benchCrc32NoIntrinsic 31 avgt 5 126.806 ±
7.301 ns/op
ChecksumBench.benchCrc32NoIntrinsic 131 avgt 5 197.751 ±
5.066 ns/op
ChecksumBench.benchCrc32NoIntrinsic 517 avgt 5 525.592 ±
10.452 ns/op
ChecksumBench.benchCrc32NoIntrinsic 2041 avgt 5 1801.082 ±
26.681 ns/op
ChecksumBench.benchCrc32c 31 avgt 5 22.610 ±
2.567 ns/op
ChecksumBench.benchCrc32c 131 avgt 5 38.121 ±
1.595 ns/op
ChecksumBench.benchCrc32c 517 avgt 5 136.786 ±
6.541 ns/op
ChecksumBench.benchCrc32c 2041 avgt 5 140.849 ±
3.359 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 31 avgt 5 45.763 ±
8.204 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 131 avgt 5 107.238 ±
7.684 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 517 avgt 5 373.500 ±
11.936 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 2041 avgt 5 1482.141 ±
27.577 ns/op
{code}
{code:title=JDK17}
Benchmark (bufferSize) Mode Cnt Score
Error Units
ChecksumBench.benchCrc32 31 avgt 5 67.810 ±
4.189 ns/op
ChecksumBench.benchCrc32 131 avgt 5 48.181 ±
1.989 ns/op
ChecksumBench.benchCrc32 517 avgt 5 75.172 ±
3.137 ns/op
ChecksumBench.benchCrc32 2041 avgt 5 179.037 ±
8.717 ns/op
ChecksumBench.benchCrc32NoIntrinsic 31 avgt 5 178.387 ±
13.835 ns/op
ChecksumBench.benchCrc32NoIntrinsic 131 avgt 5 179.689 ±
8.844 ns/op
ChecksumBench.benchCrc32NoIntrinsic 517 avgt 5 514.768 ±
13.495 ns/op
ChecksumBench.benchCrc32NoIntrinsic 2041 avgt 5 1790.131 ±
46.712 ns/op
ChecksumBench.benchCrc32c 31 avgt 5 20.638 ±
1.711 ns/op
ChecksumBench.benchCrc32c 131 avgt 5 24.648 ±
0.635 ns/op
ChecksumBench.benchCrc32c 517 avgt 5 49.179 ±
2.683 ns/op
ChecksumBench.benchCrc32c 2041 avgt 5 112.156 ±
5.765 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 31 avgt 5 46.262 ±
5.238 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 131 avgt 5 119.238 ±
2.711 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 517 avgt 5 432.203 ±
170.573 ns/op
ChecksumBench.benchCrc32cNoIntrinsic 2041 avgt 5 1549.072 ±
69.633 ns/op
{code}
> CRC32 is inefficient on x86
> ---------------------------
>
> Key: CASSANDRA-16360
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16360
> Project: Cassandra
> Issue Type: Improvement
> Components: Messaging/Client
> Reporter: Avi Kivity
> Assignee: Maxim Muzafarov
> Priority: Normal
> Labels: protocolv6
> Fix For: 5.x
>
>
> The client/server protocol specifies CRC24 and CRC32 as the checksum
> algorithm (cql_protocol_V5_framing.asc). Those however are expensive to
> compute; this affects both the client and the server.
>
> A better checksum algorithm is CRC32C, which has hardware support on x86 (as
> well as other modern architectures).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
