[
https://issues.apache.org/jira/browse/CASSANDRA-1972?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12980841#action_12980841
]
Peter Schuller commented on CASSANDRA-1972:
-------------------------------------------
I agree in general, unless the intent is to make out-of-the-box stress testing
on memory-bound workloads optimized.
In particular I agree about being queued in the kernel, and even further down
on devices or RAID controllers. Propagating outstanding I/O requests all the
way down, within reason, is key to saturating the capacity of underlying
spindles.
> Default concurrency values are improperly proportioned
> ------------------------------------------------------
>
> Key: CASSANDRA-1972
> URL: https://issues.apache.org/jira/browse/CASSANDRA-1972
> Project: Cassandra
> Issue Type: Bug
> Reporter: Stu Hood
> Assignee: Stu Hood
> Priority: Minor
> Fix For: 0.7.1
>
>
> The "default"/"suggested" {{concurrent_reads}} value is much too low. It
> assumes that CPU will be the bottleneck, rather than IO, and for most
> deployments, this will not be the case. Additionally it is better to be
> queued for IO in the kernel or on your device than in user space, because the
> former work to optimize queue order.
> Additionally, reads are much cheaper than writes in terms of CPU time (since
> writes can experience contention due to retries), so while
> {{concurrent_writes}} should probably factor in the number of cores on the
> machine, {{concurrent_reads}} should probably be calculated purely by number
> of spindles.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
