[
https://issues.apache.org/jira/browse/HBASE-15158?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15136640#comment-15136640
]
stack commented on HBASE-15158:
-------------------------------
I compared before and after using IncrementPerformanceTest
h2. UNPATCHED
{code}
2016-02-07 22:41:51,312 INFO [main] hbase.IncrementPerformanceTest:
75th=4097315.0, 95th=6597782.0, 99th=7659555.0
Performance counter stats for './hbase/bin/hbase --config
/home/stack/conf_hbase org.apache.hadoop.hbase.IncrementPerformanceTest':
157419.795645 task-clock # 2.113 CPUs utilized
2,582,885 context-switches # 0.016 M/sec
291,467 cpu-migrations # 0.002 M/sec
132,919 page-faults # 0.844 K/sec
195,230,392,765 cycles # 1.240 GHz
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
109,736,031,635 instructions # 0.56 insns per cycle
20,548,309,471 branches # 130.532 M/sec
679,777,380 branch-misses # 3.31% of all branches
74.495320701 seconds time elapsed
{code}
h2. PATCHED
{code}
2016-02-07 22:30:32,252 INFO [main] hbase.IncrementPerformanceTest:
75th=2979505.0, 95th=4664926.0, 99th=5463442.0
Performance counter stats for './hbase/bin/hbase --config
/home/stack/conf_hbase org.apache.hadoop.hbase.IncrementPerformanceTest':
157569.138177 task-clock # 2.784 CPUs utilized
2,588,954 context-switches # 0.016 M/sec
345,208 cpu-migrations # 0.002 M/sec
148,887 page-faults # 0.945 K/sec
200,320,654,874 cycles # 1.271 GHz
<not supported> stalled-cycles-frontend
<not supported> stalled-cycles-backend
118,641,158,871 instructions # 0.59 insns per cycle
22,289,641,614 branches # 141.459 M/sec
610,221,957 branch-misses # 2.74% of all branches
56.603756542 seconds time elapsed
{code}
Its 80 threads each doing 10k increments. Seems like patched does better.
> Change order in which we do write pipeline operations; do all under row locks!
> ------------------------------------------------------------------------------
>
> Key: HBASE-15158
> URL: https://issues.apache.org/jira/browse/HBASE-15158
> Project: HBase
> Issue Type: Sub-task
> Components: Performance
> Reporter: stack
> Assignee: stack
> Fix For: 2.0.0
>
> Attachments: 15158.patch, 15158v2.patch, 15158v3.patch,
> 15158v4.patch, 15158v4.patch, measurements.tgz
>
>
> Change how we do our write pipeline. I want to do all write pipeline ops
> under row lock so I lean on this fact fixing performance regression in
> check-and-set type operations like increment, append, and checkAnd* (see
> sibling issue HBASE-15082).
> To be specific, we write like this now:
> {code}
> # take rowlock
> # start mvcc
> # append to WAL
> # add to memstore
> # let go of rowlock
> # sync WAL
> # in case of error: rollback memstore
> {code}
> Instead, write like this:
> {code}
> # take rowlock
> # start mvcc
> # append to WAL
> # sync WAL
> # add to memstore
> # let go of rowlock
> ... no need to do rollback.
> {code}
> The old ordering was put in place because it got better performance in a time
> when WAL was different and before row locks were read/write (HBASE-12751).
> Testing in branch-1 shows that a reordering and skipping mvcc waits gets us
> back to the performance we had before we unified mvcc and sequenceid
> (HBASE-8763). Tests in HBASE-15046 show that at the macro level using our
> usual perf tools, reordering pipeline seems to cause no slowdown (see
> HBASE-15046). A rough compare of increments with reordered write pipeline
> seems to have us getting back a bunch of our performance (see tail of
> https://issues.apache.org/jira/browse/HBASE-15082?focusedCommentId=15111703&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15111703
> and subsequent comment).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
