[ https://issues.apache.org/jira/browse/HDFS-2379?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13117073#comment-13117073 ]
Todd Lipcon commented on HDFS-2379: ----------------------------------- We have some customers who have lots of small blocks (unfortunately they don't make good use of HAR). So, a single drive may have 400k+ blocks. When there's a lot of page cache pressure and the dentry/inode caches get pushed out, we're seeing it take several minutes per drive to do the scan. I've been experimenting with tuning /proc/sys/vm/vfs_cache_pressure which seems to help some, but even still it's taking many seconds when under lots of load. (eg in the middle of a terasort) It was a little tricky to get right, but this patch includes a "sanity check" mode which I used to catch several bugs. I think given that, today, we don't even properly synchronize it, the chance that this introduces more bugs is low. Still, I'm running some continuous cluster tests with this patch -- HBase write workloads with block report interval 90s. This shuffles through a lot of blocks quickly and helped me find some issues while working on the patch. > 0.20: Allow block reports to proceed without holding FSDataset lock > ------------------------------------------------------------------- > > Key: HDFS-2379 > URL: https://issues.apache.org/jira/browse/HDFS-2379 > Project: Hadoop HDFS > Issue Type: Bug > Components: data-node > Affects Versions: 0.20.206.0 > Reporter: Todd Lipcon > Priority: Critical > Attachments: hdfs-2379.txt, hdfs-2379.txt > > > As disks are getting larger and more plentiful, we're seeing DNs with > multiple millions of blocks on a single machine. When page cache space is > tight, block reports can take multiple minutes to generate. Currently, during > the scanning of the data directories to generate a report, the FSVolumeSet > lock is held. This causes writes and reads to block, timeout, etc, causing > big problems especially for clients like HBase. > This JIRA is to explore some of the ideas originally discussed in HADOOP-4584 > for the 0.20.20x series. -- 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