[jira] [Updated] (HBASE-12719) Add test WAL provider to quantify FSHLog overhead in the absence of HDFS.

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

     [ https://issues.apache.org/jira/browse/HBASE-12719?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Sean Busbey updated HBASE-12719:
--------------------------------
    Attachment: HBASE-12719.1.patch.txt
                HBASE-12719_comparison_chart.tiff

attaching an implementation as well as a chart showing the perf change going from disabled WAL to various amounts of HDFS interactions.

if you look at "HBASE-12719_comparison_chart" you can see the disabled WAL (which is all noops) in red at the top. The cyan line is FSHLog but w/o any HDFS interactions. The default FSHLog and the new test driver but with all interactions enabled are at the bottom; they've drawn over each other so only one is apparent.

> Add test WAL provider to quantify FSHLog overhead in the absence of HDFS.
> -------------------------------------------------------------------------
>
>                 Key: HBASE-12719
>                 URL: https://issues.apache.org/jira/browse/HBASE-12719
>             Project: HBase
>          Issue Type: Improvement
>          Components: test, wal
>            Reporter: Sean Busbey
>            Assignee: Sean Busbey
>            Priority: Minor
>             Fix For: 2.0.0, 1.1.0
>
>         Attachments: HBASE-12719.1.patch.txt, HBASE-12719_comparison_chart.tiff
>
>
> Discussion in HBASE-5699 included showing the max throughput for our WAL benchmark (using the DisabledWALProvider) compared to when we are actually dealing with synchronization and talking with HDFS.
> What we can't see in that comparison is how much of the (sizable) gap is due to the coordination done in FSHLog to deal with multi-threading and how much of it is due to talking to HDFS.
> Make a test-only provider that we can use to isolate the cost for HDFS appends, HDFS flush, and file rolling.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)