[ https://issues.apache.org/jira/browse/CASSANDRA-8160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14334197#comment-14334197 ]
Matt Stump commented on CASSANDRA-8160: --------------------------------------- Can it be done in a platform independent way that doesn't involve actually reading the files via the JVM and materializing objects on the heap? I want to avoid exacerbating our parnew collection problems. My only other concern is that when working with a large customer they said that they had written a series of sstables to disk, and hadn't rebooted the node. Again, the dataset per node was smaller than the memory available for buffer cache. After performing the /dev/null trick they saw an improvement in latencies, which should only be true for sstables which were written prior to a reboot. I was kept at a distance so I couldn't verify whether or not the files were present in the buffer cache prior to the command being run. The implication would be that the OS (RHEL 6.4) wasn't caching newly written files despite an excess of available memory. This was with DSE > 4.0 so the 1.2 bug where we were gratuitously invalidating the buffer cache shouldn't have been in effect. I'm not stating that we should necessarily design around this edge case, but we should be aware of it's possibility and try to replicate it in testing. > CF level option to call posix_fadvise for sstables on creation and startup > -------------------------------------------------------------------------- > > Key: CASSANDRA-8160 > URL: https://issues.apache.org/jira/browse/CASSANDRA-8160 > Project: Cassandra > Issue Type: New Feature > Components: Core > Reporter: Matt Stump > Assignee: Branimir Lambov > Priority: Minor > Fix For: 2.1.4 > > Attachments: trunk-8160.txt > > > We should have a CF level configuration with will result in posix_fadvise > being called for sstables for that CF. It should be called on node startup > and for new sstables. This should be configurable per CF to allow for some > CFs to be prioritized above others. Not sure if we should use > POSIX_FADV_SEQUENTIAL or POSIX_FADV_WILLNEED. -- This message was sent by Atlassian JIRA (v6.3.4#6332)