[
https://issues.apache.org/jira/browse/CASSANDRA-8457?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14260565#comment-14260565
]
Ariel Weisberg commented on CASSANDRA-8457:
-------------------------------------------
I reran this with placement groups. Took a while to get cluster setup scripted.
Unfortunately I couldn't go past 20 nodes due to a restriction on # of spot
instances. Might be able to start regular instances in the same placement
group. Still working on it.
I upped the number of servers to 14 thinking the clients weren't heavily loaded
since I got more throughput per client with only 3 servers and there is not a
lot of load on them. This is going from 9 to 14 servers. Throughput was worse
so something is wrong. This is a different kernel, I went from 12.04 with the
3.2 kernel to 14.04 running 3.13 so more then one variable has changed.
I wanted to add more clients to see if that would get the throughput up.
Something is very fishy. Clients claim they are fully connected to the cluster.
They print "New Cassandra host /172.31.56.45:9042 added" for each server node.
||Run|trunk|modified||
|#1|101,395|114,981|
|#2|123,916|98,308|
|#3|117,831|109,442|
|#4|120,267|111,571|
Going to try and find out why adding capacity did not yield more throughput so
I can start running valid benchmarks.
> nio MessagingService
> --------------------
>
> Key: CASSANDRA-8457
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8457
> Project: Cassandra
> Issue Type: New Feature
> Components: Core
> Reporter: Jonathan Ellis
> Assignee: Ariel Weisberg
> Labels: performance
> Fix For: 3.0
>
>
> Thread-per-peer (actually two each incoming and outbound) is a big
> contributor to context switching, especially for larger clusters. Let's look
> at switching to nio, possibly via Netty.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
