[
https://issues.apache.org/jira/browse/HBASE-12213?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14612900#comment-14612900
]
ramkrishna.s.vasudevan commented on HBASE-12213:
------------------------------------------------
Did some perf testing using the Performance evaluation tool.
For range scan there is no difference between with and without patch
ScanRange10000 rows=10000 with 25 threads, filterAll = true
without patch
{code}
2015-07-03 17:06:35,332 INFO [main] hbase.PerformanceEvaluation:
[RandomScanWithRange10000Test] Summary of timings (ms): [573388, 574365,
573071, 571351, 572586, 572020, 572732, 573370, 573627, 573146, 573245, 574017,
573049, 574283, 571470, 574312, 574539, 574515, 571540, 574108, 573530, 574177,
574143, 572927, 574541]
2015-07-03 17:06:35,333 INFO [main] hbase.PerformanceEvaluation:
[RandomScanWithRange10000Test] Min: 571351ms Max: 574541ms Avg:
573362ms
{code}
with patch
{code}
2015-07-03 16:49:57,711 INFO [main] hbase.PerformanceEvaluation:
[RandomScanWithRange10000Test] Summary of timings (ms): [572276, 569970,
572610, 571517, 572027, 570168, 570932, 572951, 570492, 572410, 572037, 571495,
572238, 572055, 572114, 571142, 572500, 569813, 572383, 572223, 571191, 572197,
571473, 571235, 569209]
2015-07-03 16:49:57,715 INFO [main] hbase.PerformanceEvaluation:
[RandomScanWithRange10000Test] Min: 569209ms Max: 572951ms Avg:
571546ms
{code}
But when we do gets which involves lot of blockSeek code
randomReads = 10 threads, multiGets = 100, filterAll = true
withoutpatch
{code}
2015-07-03 17:13:37,172 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Summary of timings (ms): [158292, 158191, 157503, 158375,
158182, 158901, 157680, 158583, 158388, 158279]
2015-07-03 17:13:37,173 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Min: 157503ms Max: 158901ms Avg: 158237ms
2015-07-03 17:33:36,854 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Summary of timings (ms): [163757, 163478, 163175, 163876,
163490, 163187, 163813, 163937, 162839, 163398]
2015-07-03 17:33:36,855 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Min: 162839ms Max: 163937ms Avg: 163495ms
2015-07-03 17:36:44,255 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Summary of timings (ms): [161414, 162715, 162266, 162409,
161884, 162161, 162235, 162077, 162852, 161575]
2015-07-03 17:36:44,256 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Min: 161414ms Max: 162852ms Avg: 162158ms
{code}
with patch
{code}
2015-07-03 17:26:02,496 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Summary of timings (ms): [150311, 150721, 150350, 149602,
150339, 150593, 151117, 149046, 150403, 150087]
2015-07-03 17:26:02,497 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Min: 149046ms Max: 151117ms Avg: 150256ms
2015-07-03 17:43:52,968 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Summary of timings (ms): [148812, 148077, 148482, 148160,
148479, 147835, 148461, 147683, 148499, 149299]
2015-07-03 17:43:52,970 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Min: 147683ms Max: 149299ms Avg: 148378ms
2015-07-03 17:46:30,247 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Summary of timings (ms): [149254, 148839, 148658, 148982,
148443, 147960, 148590, 148508, 148855, 148767]
2015-07-03 17:46:30,248 INFO [main] hbase.PerformanceEvaluation:
[RandomReadTest] Min: 147960ms Max: 149254ms Avg: 148685ms
{code}
We see significant gain. Currently note that we are copying the offheap
buckets to the onheap single BB. The blockSeek positional based reads and some
optimization in the KeyOnlyKV is the reason for this performance gain. Even
without MBB we should be able to see the same gain because now the MBB is going
to act as a wrapper over the single BB that was formed from the underlying
offheap buckets.
> HFileBlock backed by Array of ByteBuffers
> -----------------------------------------
>
> Key: HBASE-12213
> URL: https://issues.apache.org/jira/browse/HBASE-12213
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver, Scanners
> Reporter: Anoop Sam John
> Assignee: ramkrishna.s.vasudevan
> Attachments: HBASE-12213_1.patch
>
>
> In L2 cache (offheap) an HFile block might have been cached into multiple
> chunks of buffers. If HFileBlock need single BB, we will end up in recreation
> of bigger BB and copying. Instead we can make HFileBlock to serve data from
> an array of BBs.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
