enixon commented on issue #851: ZOOKEEPER-3311: Allow a delay to the transaction log flush URL: https://github.com/apache/zookeeper/pull/851#issuecomment-489862774 Internally, we ran a test that starts with no flush delay configured (eg. the default which has no flush delay and a max batch size of 1,000) and after a few minutes uses JMX to change the settings. Both the flush delay and max queue poll time are set to 20ms, while the max batch size is set to 0 (which disables flushing based on batch size and only flushes based on time). The test uses 10 client machines that each do 500 writes/s for a total load of 5000 write/s to the ensemble, each write creating a unique path with a 16-byte random value. After the flush delay is enabled, we see differences in various metrics. The number of fsyncs per second drops from the mid-hundreds before the delay to 50 after (which is expected, 1000ms / 20ms = 50). Other interesting metrics are average fsynctime and minimum fsynctime which both show a marked improvement in per-fsync time as a consequence of doing fewer a second.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
