Check the logs for compaction.  CPU will go up while one node may be compacting 
and the other node is not compacting yet.  Compaction can also last quite a 
long time in some cases.  Also, you can do a jstack –l {pid} > thread.txt to do 
a thread dump to see what that node is doing and compare it with other nodes.

I am actually in the process of doing quite a bit of performance analysis 
myself and I know I need to go back and tweak the astyanax client somehow as 
well as I found I can run a test of running jmap heap dump on node 1 in my 
cluster and our website sometimes slows down(ie. I keep refreshing the pages) 
and I feel astyanax is going back to node 1 too often which results in a 10 
second page load vs. when it hits nodes that are fine, it loads in 300ms.

We have been running live for a while and get various percentage bumps on some 
nodes vs. others but mostly has been compaction running at different times for 
quite a while.

Dean

From: Marcelo Elias Del Valle <mvall...@gmail.com<mailto:mvall...@gmail.com>>
Reply-To: "user@cassandra.apache.org<mailto:user@cassandra.apache.org>" 
<user@cassandra.apache.org<mailto:user@cassandra.apache.org>>
Date: Wednesday, February 20, 2013 3:53 PM
To: "user@cassandra.apache.org<mailto:user@cassandra.apache.org>" 
<user@cassandra.apache.org<mailto:user@cassandra.apache.org>>
Subject: load on cluster nodes

Hello,

     I am writing a lot of data on a cassandra cluster with 3 nodes. All my 
three nodes are seeds, when I create the astyanax connection pool I specify 
ip1,ip2,ip3 as seeds and so in the casandra.yaml config file.
      However, if I look at the server load when I am bombarding it with data, 
I can see the first server is clearly being more demanded (graph here: 
http://mvalle.com/downloads/serverload1.png) . I am writting data in sequence, 
generating UUIDs for column families based on timestamp. Also, I am using 
default partitioning (cassandra 1.1.9).
      Any hint why this is happening and how to solve it? I would expect server 
loads to be more equal along the cluster...

Best regards,
--
Marcelo Elias Del Valle
http://mvalle.com - @mvallebr

Reply via email to