[ https://issues.apache.org/jira/browse/HADOOP-13047?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15256525#comment-15256525 ]
Steve Loughran commented on HADOOP-13047: ----------------------------------------- HADOOP-13028 has an attachment which compares read bandwidths to open time...from the numbers locally to the US, I'm seeing a forward skip of < 800K would be faster than closing and reopening the connection —and that's ignoring the effects of TCP slow start on followon reads {code} 2016-04-25 17:01:09,584 INFO scale.S3AScaleTestBase (S3AScaleTestBase.java:end(181)) - Duration of Open stream: 164,965,000 nS 2016-04-25 17:01:13,613 INFO scale.S3AScaleTestBase (S3AScaleTestBase.java:end(181)) - Duration of Time to read 20405780 bytes: 4,029,203,000 nS 2016-04-25 17:01:13,614 INFO scale.TestS3AInputStreamPerformance (TestS3AInputStreamPerformance.java:bandwidth(135)) - Bandwidth = 5.064471 MB/S 2016-04-25 17:01:13,614 INFO scale.TestS3AInputStreamPerformance (TestS3AInputStreamPerformance.java:testTimeToOpenAndReadWholeFileByByte(97)) - An open() call has the equivalent duration of reading 837,385 bytes {code} > S3a Forward seek in stream length to be configurable > ---------------------------------------------------- > > Key: HADOOP-13047 > URL: https://issues.apache.org/jira/browse/HADOOP-13047 > Project: Hadoop Common > Issue Type: Sub-task > Components: fs/s3 > Affects Versions: 2.8.0 > Reporter: Steve Loughran > Attachments: HADOOP-13047.WIP.2.patch, HADOOP-13047.WIP.patch > > > Even with lazy seek, tests can show that sometimes a short-distance forward > seek is triggering a close + reopen, because the threshold for the seek is > simply available bytes in the inner stream. > A configurable threshold would allow data to be read and discarded before > that seek. This should be beneficial over long-haul networks as the time to > set up the TCP channel is high, and TCP-slow-start means that the ramp up of > bandwidth is slow. In such deployments, it will better to read forward than > re-open, though the exact "best" number will vary with client and endpoint. -- This message was sent by Atlassian JIRA (v6.3.4#6332)