[
https://issues.apache.org/jira/browse/HBASE-12040?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14152648#comment-14152648
]
Nick Dimiduk commented on HBASE-12040:
--------------------------------------
Ran the following on single node with separate processes for NN, DN, ZK,
HMaster, RS. HDFS 2.4.0, SCR enabled, 5 DFS drives. For branch-1 there's no RS
process. Configured with 20g heap, LRU blockcache, hfile.block.cache.size=0.6,
hbase.block.data.cachecompressed is not specified. I use scanRange100 to be
consistent with the spirit of the original report, but using a much larger data
size.
Not an explanation of why, but here's what I came out with. Number in parens is
pct change vs 0.98.6.1. This implies we've leaked a bit of perf on 0.98 branch
since 0.98.6.1 release and it's gotten worse on branch-1.
|| build | overall (ms) | avg (μs) | 95% (μs) ||
| 0.98.6.1 | 164174 | 15542 | 29649 |
| b885130~1 | 195675 ({color:red}19.19%{color}) | 18547
({color:red}19.33%{color}) | 33337 ({color:red}12.44%{color}) |
| b885130 | 192937 ({color:red}17.52%{color}) | 18290
({color:red}17.68%{color}) | 32750 ({color:red}10.46%{color}) |
| 0.99.0RC1 | 201521 ({color:red}22.75%{color}) | 19104
({color:red}22.92%{color}) | 33768 ({color:red}13.89%{color}) |
Data generated on with RS running 0.98.6.1, using a client built from master
(latest PE features).
{noformat}
$ hbase org.apache.hadoop.hbase.PerformanceEvaluation --nomapred
--compress=SNAPPY --presplit=10 --size=100 sequentialWrite 1
{noformat}
Here's the rough script I ran to get each value:
1. deploy target version
2. drop buffer caches
{noformat}
$ sync; echo 3 | sudo tee /proc/sys/vm/drop_caches
3
{noformat}
3. run test
{noformat}
$ hbase org.apache.hadoop.hbase.PerformanceEvaluation --nomapred --size=100
--sampleRate=0.0001 scanRange100 1
{noformat}
0.98.6.1 and 0.99.0RC1 are from official releases; other two are built locally
using
{noformat}
$ git checkout XXX
$ bash -x ./dev-support/generate-hadoopX-poms.sh 0.98.7-SNAPSHOT
0.98.7-hadoop2-SNAPSHOT
$ mvn -f pom.xml.hadoop2 clean install assembly:single -DskipTests
{noformat}
[~jmspaggi] I think you've found something.
> Performances issues with FilteredScanTest
> ------------------------------------------
>
> Key: HBASE-12040
> URL: https://issues.apache.org/jira/browse/HBASE-12040
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.99.1
> Reporter: Jean-Marc Spaggiari
> Priority: Critical
> Attachments: at-HBASE-11331.html, pre-HBASE-11331.html,
> ycsb-0.98.6.1-vs-0.98.7.pdf
>
>
> While testing 0.99.0RC1 release performances, compared to 0.98.6, figured
> that:
> - FilteredScanTest is 100 times slower;
> - RandomReadTest is 1.5 times slower;
> - RandomSeekScanTest is 3.2 times slower;
> - RandomScanWithRange10Test is 1,2 times slower;
> - RandomScanWithRange100Test is 1,3 times slower;
> - RandomScanWithRange1000Test is 4 times slower;
> - SequentialReadTest is 1,7 times slower;
> - SequentialWriteTest is just a bit faster;
> - RandomWriteTest is just a bit faster;
> - GaussianRandomReadBenchmark is just a beat slower;
> - SequentialReadBenchmark is 1,1 times slower;
> - SequentialWriteBenchmark is 1,1 times slower;
> - UniformRandomReadBenchmark crashed;
> - UniformRandomSmallScan is 1,3 times slower.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
