[ https://issues.apache.org/jira/browse/CASSANDRA-6357?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13881433#comment-13881433 ]
dan jatnieks commented on CASSANDRA-6357: ----------------------------------------- I did some stress write testing on a single node physical machine with multiple rotating disks. The machine was a quad-code Intel Xeon E3-1230, 32Gb memory, and 4 disks (2Tb, 5400 rpm). Two scenarios were measured using stress-write for 10 million records (cassandra-stress --num-keys 10000000 --columns=50 --operation=INSERT): Base scenario: * O/S on separate device * Commitlog on separate device * Data and flush directories on the same device Flush scenario: * O/S on separate device * Commitlog on separate device * Data directory on separate device * Flush directory on separate device The result of splitting the flush directory to another device on the stress-write results was: * an elapsed time improvement of about 57% * a much lower, and consistent, 99.9th percentile latency (3308 ms vs 15963 ms) See attached graph: [Stress Write Latency 99.9th Percentile|^c6357-stress-write-latency-99th-1.png] > Flush memtables to separate directory > ------------------------------------- > > Key: CASSANDRA-6357 > URL: https://issues.apache.org/jira/browse/CASSANDRA-6357 > Project: Cassandra > Issue Type: Improvement > Components: Core > Reporter: Patrick McFadin > Assignee: Jonathan Ellis > Attachments: 6357.txt > > > Flush writers are a critical element for keeping a node healthy. When several > compactions run on systems with low performing data directories, IO becomes a > premium. Once the disk subsystem is saturated, write IO is blocked which will > cause flush writer threads to backup. Since memtables are large blocks of > memory in the JVM, too much blocking can cause excessive GC over time > degrading performance. In the worst case causing an OOM. > Since compaction is running on the data directories. My proposal is to create > a separate directory for flushing memtables. Potentially we can use the same > methodology of keeping the commit log separate and minimize disk contention > against the critical function of the flushwriter. -- This message was sent by Atlassian JIRA (v6.1.5#6160)