[ https://issues.apache.org/jira/browse/HDFS-9666?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15110188#comment-15110188 ]
Yu Li commented on HDFS-9666: ----------------------------- bq. However it looked like the benefits of reading from remote RAM were canceled by the RPC overhead, as compared to short-circuit reads from local disk Agreed this is true for most *common* case. However, since SATA has much poor io performance than SSD/RAM, reading from remote SSD/RAM is useful to reduce spike in the system, or say it's good for reducing the Max latency rather than Avg. And since there's a switch to turn on/off the feature, user could choose to use it or not according to different scenarios. > Enable hdfs-client to read even remote SSD/RAM prior to local disk replica to > improve random read > ------------------------------------------------------------------------------------------------- > > Key: HDFS-9666 > URL: https://issues.apache.org/jira/browse/HDFS-9666 > Project: Hadoop HDFS > Issue Type: Improvement > Components: hdfs-client > Affects Versions: 2.6.0, 2.7.0 > Reporter: ade > Assignee: ade > Fix For: 2.7.2 > > Attachments: HDFS-9666.0.patch > > > We want to improve random read performance of HDFS for HBase, so enabled the > heterogeneous storage in our cluster. But there are only ~50% of datanode & > regionserver hosts with SSD. we can set hfile with only ONE_SSD not ALL_SSD > storagepolicy and the regionserver on none-SSD host can only read the local > disk replica . So we developed this feature in hdfs client to read even > remote SSD/RAM prior to local disk replica. -- This message was sent by Atlassian JIRA (v6.3.4#6332)