[
https://issues.apache.org/jira/browse/CASSANDRA-16342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17271892#comment-17271892
]
Yifan Cai commented on CASSANDRA-16342:
---------------------------------------
[~marcuse], the difference of the live sstables count could be related with
altering the compression `chunk_length_in_kb` from 64 to 16 in the 3.0 cluster.
The default values of `chunk_length_in_kb` are different in 3.0 and 4.0
clusters. They are 64 and 16 respectively. Therefore, the parameter was altered
to match with what was in 4.0 after data population, as I only noticed it by
then. When the chunk length is 64kb, for the populated data, there are less
sstables.
The other 30 v 40 perf test runs were updated to create table with 16kb chunk
length explicitly after the first run. So the other results are not affected.
Attaching the link to revised result report from the latest run:
[https://github.com/yifan-c/CASSANDRA-15581-COMPACTION-TEST/blob/main/CASSANDRA-16342-16343/%5BRevised%5D%20Compaction%20Perf%20Comparison%20Between%203.0%20and%204.0%20(STCS%20and%20LCS).pdf]
Under the same load, the LCS compaction in 4.0 has better performance. The
compaction in 3.0 cluster is lagging behind as the number of the unleveled
sstables increases during the test. Meanwhile, the 4.0 cluster holds up to the
load. The number of pending tasks, the compaction throughput and the unleveled
sstables count remain steady throughout the test.
The number of the live sstables are similar for both clusters.
> LCS steady state load of 4.0 vs. 3.x performance test
> -----------------------------------------------------
>
> Key: CASSANDRA-16342
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16342
> Project: Cassandra
> Issue Type: Sub-task
> Components: Test/benchmark
> Reporter: Yifan Cai
> Assignee: Yifan Cai
> Priority: Normal
> Attachments: 30_40_cluster_load.png
>
>
> The testing cluster should be pre-populated with ~200GB data in each node.
> Run the steady state workload to compare the read, write and compaction
> performance between baseline (3.x cluster) and 4.0 cluster.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
