cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Philip Thompson (JIRA)" <>
Subject [jira] [Commented] (CASSANDRA-9504) Odd performance numbers comparing 2.0.15 and 2.1+
Date Fri, 29 May 2015 21:00:19 GMT


Philip Thompson commented on CASSANDRA-9504:

After discussion with [~brandon.williams] and [~benedict], it appears this is expected. 2.0
is apparently not respecting the batch window, a bug that was fixed for 2.1+. To get comparable
performance out of later versions, the batch_window_in_ms was tuned down to 1-5ms and concurrent_writers
was increased. Here is a graph showing the change:

It seems that we should consider lowering the default batch window, or removing the window
altogether, since the default performance is drastically worse. Changing this to Improvement,
to reflect.

> Odd performance numbers comparing 2.0.15 and 2.1+
> -------------------------------------------------
>                 Key: CASSANDRA-9504
>                 URL:
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Philip Thompson
>             Fix For: 3.x, 2.2.x
>         Attachments: jstack.tar.gz
> I've been doing some basic performance testing of batch commitlog on 2.0.15, 2.1-head,
and 2.2.-head. I have been seeing very odd results where 2.0.15 performance is incredibly
high, suspiciously so, when compared to 2.1 or 2.2.
> Example graph:
> I'm using the following yaml options:
> {code}
> commitlog_sync: batch
> commitlog_sync_batch_window_in_ms: 50
> commitlog_sync_period_in_ms: null
> concurrent_writes: 64
> {code}
> I am able to reproduce this on two separate clusters, physical hardware, with SSDs. I
have not been able to reproduce this on gce, or a physical cluster I have access to with HDD.
I can reproduce this consistently, with different stress options, and with higher data load
on the nodes. I can also reproduce this myself on these machines, without using the C* perf
tool to set up the cluster and perform the operations.
> When running 2.0.15 batch commitlog vs periodic, I do see a difference as expected:
> I am using the newest stress code on trunk. What should I be looking for to explain the
anomalous performance here?

This message was sent by Atlassian JIRA

View raw message