[ https://issues.apache.org/jira/browse/HDFS-3697?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13421075#comment-13421075 ]
Brandon Li commented on HDFS-3697: ---------------------------------- Hey Todd, I've seen some performance improvement in my branch-1 test with the fadvise support enabled. I think should commit it to branch-1. The thing I am not sure is the readahead step size since I didn't try out with enough workloads and step sizes in my tests. Given 4MB is good in your experiments and it's also configurable, we can start with it. I just feel it's not trivial sometimes to guess out a good step size beforehand. :-) If it's possible in the future, we might want to make readahead adjust its step size based on observed access pattern(however, tracking multiple parallel streams' access pattern can be same challenging :-( ). > Enable fadvise readahead by default > ----------------------------------- > > Key: HDFS-3697 > URL: https://issues.apache.org/jira/browse/HDFS-3697 > Project: Hadoop HDFS > Issue Type: Improvement > Components: data-node, performance > Affects Versions: 3.0.0, 2.2.0-alpha > Reporter: Todd Lipcon > Assignee: Todd Lipcon > Priority: Minor > Fix For: 3.0.0, 2.2.0-alpha > > Attachments: hdfs-3697-branch-1.txt, hdfs-3697.txt, hdfs-3697.txt > > > The fadvise features have been implemented for some time, and we've enabled > them in production at a lot of customer sites without difficulty. I'd like to > enable the readahead feature by default in future versions so that users get > this benefit without any manual configuration required. > The other fadvise features seem to be more workload-dependent and need > further testing before enabling by default. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira