[
https://issues.apache.org/jira/browse/CASSANDRA-5530?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13662464#comment-13662464
]
Pavel Yaskevich commented on CASSANDRA-5530:
--------------------------------------------
So there is a fundamental problem with CustomTHsHaServer and that is why it
runs so fast on stress benchmarks, CustomTHsHaServer.requestInvoke doesn't use
an invoker queue given to it and instead just executes requests inline in the
selector thread context. So for any heavy request (read/write) it actually
would become even worse than "sync" server. On the other hand CASSANDRA-5582
utilizes the invocation queue correctly, that is why (because of additional
context-switch and because it wasn't using custom ExecutorService) it shows
worse results with stress tests.
> Switch from THSHAServer to TThreadedSelectorServer
> --------------------------------------------------
>
> Key: CASSANDRA-5530
> URL: https://issues.apache.org/jira/browse/CASSANDRA-5530
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: T Jake Luciani
> Assignee: T Jake Luciani
> Fix For: 2.0
>
> Attachments: 5530.txt
>
>
> TThreadedSelectorServer is new in Thrift 0.9. It builds on HSHA by allowing
> for a set of threads for IO and a set work request processing...
> I've attached the performance numbers below. It's a lot closer to
> TThreadedServer.
> ThreadedServer (Default)
> {code}
> Write Averages from the middle 80% of values:
> interval_op_rate : 14811
> interval_key_rate : 14811
> latency median : 1.7
> latency 95th percentile : 5.3
> latency 99.9th percentile : 142.6
> Total operation time : 00:01:16
> END
> Read Averages from the middle 80% of values:
> interval_op_rate : 16898
> interval_key_rate : 16898
> latency median : 2.2
> latency 95th percentile : 8.5
> latency 99.9th percentile : 165.7
> Total operation time : 00:01:05
> END
> {code}
> HSHA (CURRENT)
> {code}
> Write Averages from the middle 80% of values:
> interval_op_rate : 8939
> interval_key_rate : 8939
> latency median : 5.0
> latency 95th percentile : 10.1
> latency 99.9th percentile : 105.4
> Total operation time : 00:01:56
> END
> Read Averages from the middle 80% of values:
> interval_op_rate : 9608
> interval_key_rate : 9608
> latency median : 5.1
> latency 95th percentile : 7.7
> latency 99.9th percentile : 51.6
> Total operation time : 00:01:49
> END
> {code}
> TThreadedSelectorServer (NEW)
> {code}
> Write Averages from the middle 80% of values:
> interval_op_rate : 11640
> interval_key_rate : 11640
> latency median : 3.1
> latency 95th percentile : 10.6
> latency 99.9th percentile : 135.9
> Total operation time : 00:01:30
> END
> Read Averages from the middle 80% of values:
> interval_op_rate : 15247
> interval_key_rate : 15247
> latency median : 2.8
> latency 95th percentile : 7.1
> latency 99.9th percentile : 40.3
> Total operation time : 00:01:06
> END
> {code}
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
