[jira] [Updated] (HBASE-10094) Add batching to HLogPerformanceEvaluation

["stack (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HBASE-10094?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

stack updated HBASE-10094:
--------------------------

    Attachment: 10094v2.txt

Add some metrics on sync and rate at which we are writing.  Below is sample of what we have now going against localfs and then same passing a syncInterval flag set to 100:

org.apache.hadoop.hbase.regionserver.wal.HLogPerformanceEvaluation:
  append:
             count = 680000000
         mean rate = 67244.75 bytes/ms
     1-minute rate = 52488.70 bytes/ms
     5-minute rate = 50267.51 bytes/ms
    15-minute rate = 49882.70 bytes/ms

  sync:
             count = 1000000
         mean rate = 98.88 syncs/ms
     1-minute rate = 77.16 syncs/ms
     5-minute rate = 73.89 syncs/ms
    15-minute rate = 73.33 syncs/ms

# Every 100 appends.
org.apache.hadoop.hbase.regionserver.wal.HLogPerformanceEvaluation:
  append:
             count = 585820000
         mean rate = 85046.52 bytes/ms
     1-minute rate = 74908.80 bytes/ms
     5-minute rate = 74908.80 bytes/ms
    15-minute rate = 74908.80 bytes/ms

  sync:
             count = 8643
         mean rate = 1.25 syncs/ms
     1-minute rate = 1.11 syncs/ms
     5-minute rate = 1.11 syncs/ms
    15-minute rate = 1.11 syncs/ms


We probably want to settle somewhere at syncing about every 10ms or so.

> Add batching to HLogPerformanceEvaluation
> -----------------------------------------
>
>                 Key: HBASE-10094
>                 URL: https://issues.apache.org/jira/browse/HBASE-10094
>             Project: HBase
>          Issue Type: Sub-task
>          Components: Performance, wal
>            Reporter: stack
>            Assignee: Himanshu Vashishtha
>         Attachments: 10094v2.txt
>
>
> As Himanshu points out in the the parent issue, HLogPE is using an unorthodox API appending edits to the WAL; it is using an API that is meant for tests only that does an append immediately followed by a sync call.
> In normal deploy, WAL appends are done as a bunch of appends followed by a sync on the tail of the transaction -- not a sync per append.
> This issue is about changing HLogPE to use append and then sync.  It also adds an argument so you can specifying batching of a set of appends before  the sync is called.  The latter lets HLogPE mimic multi puts that use the minibatch... which appends, appends, appends.. and then syncs.
> Assigning to Himanshu for review.



--
This message was sent by Atlassian JIRA
(v6.1#6144)