[
https://issues.apache.org/jira/browse/HIVE-8291?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Mostafa Mokhtar updated HIVE-8291:
----------------------------------
Description:
Reading from bucketed partitioned tables has significantly higher overhead
compared to non-bucketed non-partitioned files.
50% of the time is spent in these two lines of code in
OrcInputFormate.getReader()
{code}
String txnString = conf.get(ValidTxnList.VALID_TXNS_KEY,
Long.MAX_VALUE + ":");
ValidTxnList validTxnList = new ValidTxnListImpl(txnString);
{code}
{code}
Stack Trace Sample Count Percentage(%)
hive.ql.exec.tez.MapRecordSource.pushRecord() 2,981 87.215
org.apache.tez.mapreduce.lib.MRReaderMapred.next() 2,002 58.572
mapred.split.TezGroupedSplitsInputFormat$TezGroupedSplitsRecordReader.next(Object,
Object) 2,002 58.572
mapred.split.TezGroupedSplitsInputFormat$TezGroupedSplitsRecordReader.initNextRecordReader()
1,984 58.046
hive.ql.io.HiveInputFormat.getRecordReader(InputSplit, JobConf,
Reporter) 1,983 58.016
hive.ql.io.orc.OrcInputFormat.getRecordReader(InputSplit, JobConf, Reporter)
1,891 55.325
hive.ql.io.orc.OrcInputFormat.getReader(InputSplit,
AcidInputFormat$Options) 1,723 50.41
hive.common.ValidTxnListImpl.<init>(String)
934 27.326
conf.Configuration.get(String, String) 621
18.169
{code}
Another 20% of the profile is spent in MapOperator.cleanUpInputFileChangedOp
5% the CPU in
{code}
Path onepath = normalizePath(onefile);
{code}
And
15% the CPU in
{code}
onepath.toUri().relativize(fpath.toUri()).equals(fpath.toUri());
{code}
>From the profiler
{code}
Stack Trace Sample Count Percentage(%)
org.apache.hadoop.hive.ql.exec.tez.MapRecordSource.processRow(Object) 978
28.613
org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator.process(Writable)
978 28.613
org.apache.hadoop.hive.ql.exec.Operator.cleanUpInputFileChanged() 866
25.336
org.apache.hadoop.hive.ql.exec.MapOperator.cleanUpInputFileChangedOp()
866 25.336
java.net.URI.relativize(URI) 655 19.163
java.net.URI.relativize(URI, URI) 655 19.163
java.net.URI.normalize(String) 517 15.126
java.net.URI.needsNormalization(String)
372 10.884
java.lang.String.charAt(int) 235
6.875
java.net.URI.equal(String, String) 27 0.79
java.lang.StringBuilder.toString() 1 0.029
java.lang.StringBuilder.<init>() 1 0.029
java.lang.StringBuilder.append(String) 1 0.029
org.apache.hadoop.hive.ql.exec.MapOperator.normalizePath(String) 167
4.886
org.apache.hadoop.fs.Path.<init>(String) 162 4.74
org.apache.hadoop.fs.Path.initialize(String, String, String, String) 162
4.74
org.apache.hadoop.fs.Path.normalizePath(String, String) 97 2.838
org.apache.commons.lang.StringUtils.replace(String, String, String)
97 2.838
org.apache.commons.lang.StringUtils.replace(String, String,
String, int) 97 2.838
java.lang.String.indexOf(String, int) 97 2.838
java.net.URI.<init>(String, String, String, String, String)
65 1.902
{code}
was:
Reading from bucketed partitioned tables has significantly higher overhead
compared to non-bucketed non-partitioned files.
50% of the time is spent in these two lines of code in
OrcInputFormate.getReader()
{code}
String txnString = conf.get(ValidTxnList.VALID_TXNS_KEY,
Long.MAX_VALUE + ":");
ValidTxnList validTxnList = new ValidTxnListImpl(txnString);
{code}
{code}
Stack Trace Sample Count Percentage(%)
hive.ql.exec.tez.MapRecordSource.pushRecord() 2,981 87.215
org.apache.tez.mapreduce.lib.MRReaderMapred.next() 2,002
58.572
mapred.split.TezGroupedSplitsInputFormat$TezGroupedSplitsRecordReader.next(Object,
Object) 2,002 58.572
mapred.split.TezGroupedSplitsInputFormat$TezGroupedSplitsRecordReader.initNextRecordReader()
1,984 58.046
hive.ql.io.HiveInputFormat.getRecordReader(InputSplit, JobConf, Reporter) 1,983
58.016
hive.ql.io.orc.OrcInputFormat.getRecordReader(InputSplit, JobConf, Reporter)
1,891 55.325
hive.ql.io.orc.OrcInputFormat.getReader(InputSplit, AcidInputFormat$Options)
1,723 50.41
hive.common.ValidTxnListImpl.<init>(String) 934 27.326
conf.Configuration.get(String, String)
621 18.169
{code}
> Reading from partitioned bucketed tables has high overhead, 50% of time is
> spent in OrcInputFormat.getReader
> ------------------------------------------------------------------------------------------------------------
>
> Key: HIVE-8291
> URL: https://issues.apache.org/jira/browse/HIVE-8291
> Project: Hive
> Issue Type: Bug
> Components: Query Processor
> Affects Versions: 0.14.0
> Environment: cn105
> Reporter: Mostafa Mokhtar
> Assignee: Prasanth J
> Fix For: 0.14.0
>
>
> Reading from bucketed partitioned tables has significantly higher overhead
> compared to non-bucketed non-partitioned files.
> 50% of the time is spent in these two lines of code in
> OrcInputFormate.getReader()
> {code}
> String txnString = conf.get(ValidTxnList.VALID_TXNS_KEY,
> Long.MAX_VALUE + ":");
> ValidTxnList validTxnList = new ValidTxnListImpl(txnString);
> {code}
> {code}
> Stack Trace Sample Count Percentage(%)
> hive.ql.exec.tez.MapRecordSource.pushRecord() 2,981 87.215
> org.apache.tez.mapreduce.lib.MRReaderMapred.next() 2,002 58.572
>
> mapred.split.TezGroupedSplitsInputFormat$TezGroupedSplitsRecordReader.next(Object,
> Object) 2,002 58.572
>
> mapred.split.TezGroupedSplitsInputFormat$TezGroupedSplitsRecordReader.initNextRecordReader()
> 1,984 58.046
> hive.ql.io.HiveInputFormat.getRecordReader(InputSplit, JobConf,
> Reporter) 1,983 58.016
>
> hive.ql.io.orc.OrcInputFormat.getRecordReader(InputSplit, JobConf, Reporter)
> 1,891 55.325
> hive.ql.io.orc.OrcInputFormat.getReader(InputSplit,
> AcidInputFormat$Options) 1,723 50.41
> hive.common.ValidTxnListImpl.<init>(String)
> 934 27.326
> conf.Configuration.get(String, String) 621
> 18.169
> {code}
> Another 20% of the profile is spent in MapOperator.cleanUpInputFileChangedOp
> 5% the CPU in
> {code}
> Path onepath = normalizePath(onefile);
> {code}
> And
> 15% the CPU in
> {code}
> onepath.toUri().relativize(fpath.toUri()).equals(fpath.toUri());
> {code}
> From the profiler
> {code}
> Stack Trace Sample Count Percentage(%)
> org.apache.hadoop.hive.ql.exec.tez.MapRecordSource.processRow(Object) 978
> 28.613
> org.apache.hadoop.hive.ql.exec.vector.VectorMapOperator.process(Writable)
> 978 28.613
> org.apache.hadoop.hive.ql.exec.Operator.cleanUpInputFileChanged()
> 866 25.336
>
> org.apache.hadoop.hive.ql.exec.MapOperator.cleanUpInputFileChangedOp()
> 866 25.336
> java.net.URI.relativize(URI) 655 19.163
> java.net.URI.relativize(URI, URI) 655 19.163
> java.net.URI.normalize(String) 517 15.126
> java.net.URI.needsNormalization(String)
> 372 10.884
> java.lang.String.charAt(int) 235
> 6.875
>
> java.net.URI.equal(String, String) 27 0.79
>
> java.lang.StringBuilder.toString() 1 0.029
>
> java.lang.StringBuilder.<init>() 1 0.029
>
> java.lang.StringBuilder.append(String) 1 0.029
>
> org.apache.hadoop.hive.ql.exec.MapOperator.normalizePath(String) 167
> 4.886
>
> org.apache.hadoop.fs.Path.<init>(String) 162 4.74
>
> org.apache.hadoop.fs.Path.initialize(String, String, String, String) 162
> 4.74
> org.apache.hadoop.fs.Path.normalizePath(String, String) 97 2.838
> org.apache.commons.lang.StringUtils.replace(String, String, String)
> 97 2.838
> org.apache.commons.lang.StringUtils.replace(String, String,
> String, int) 97 2.838
> java.lang.String.indexOf(String, int) 97 2.838
> java.net.URI.<init>(String, String, String, String, String)
> 65 1.902
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
