[
https://issues.apache.org/jira/browse/HIVE-9495?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14319633#comment-14319633
]
Mostafa Mokhtar commented on HIVE-9495:
---------------------------------------
[~navis]
I believe this is why this operation is slow
{code}
hashAggregations = new HashMap<KeyWrapper, AggregationBuffer[]>(256);
{code}
We should be using the estimate row count to correctly size this HashMap to
avoid excessive chaining and resizing, yet if we estimate too much this can
cause OOM.
Inserting a million rows into a hash map with an initial size of 256 is likely
to result in bad performance.
Something like this from HashTableLoader.load
{code}
Map<Integer, Long> parentKeyCounts = desc.getParentKeyCounts();
Long keyCountObj = parentKeyCounts.get(pos);
long keyCount = (keyCountObj == null) ? -1 : keyCountObj.longValue();
{code}
Ideally the decision to enable map side aggregation should be driven by CE
(cardinality estimate) and NDV.
Based on these two we can estimate how much reduction we get from the map side
aggregation, in other words if NDV = CE then skip map side aggregation.
For TPC-H each l_orderkey is repeated ~4 times, so we are better off skipping
the map side aggregation (local agg).
> Map Side aggregation affecting map performance
> ----------------------------------------------
>
> Key: HIVE-9495
> URL: https://issues.apache.org/jira/browse/HIVE-9495
> Project: Hive
> Issue Type: Bug
> Components: Query Processor
> Affects Versions: 0.14.0
> Environment: RHEL 6.4
> Hortonworks Hadoop 2.2
> Reporter: Anand Sridharan
> Attachments: HIVE-9495.1.patch.txt, profiler_screenshot.PNG
>
>
> When trying to run a simple aggregation query with hive.map.aggr=true, map
> tasks take a lot of time in Hive 0.14 as against with hive.map.aggr=false.
> e.g.
> Consider the query:
> {code}
> INSERT OVERWRITE TABLE lineitem_tgt_agg
> select alias.a0 as a0,
> alias.a2 as a1,
> alias.a1 as a2,
> alias.a3 as a3,
> alias.a4 as a4
> from (
> select alias.a0 as a0,
> SUM(alias.a1) as a1,
> SUM(alias.a2) as a2,
> SUM(alias.a3) as a3,
> SUM(alias.a4) as a4
> from (
> select lineitem_sf500.l_orderkey as a0,
> CAST(lineitem_sf500.l_quantity * lineitem_sf500.l_extendedprice * (1 -
> lineitem_sf500.l_discount) * (1 + lineitem_sf500.l_tax) as double) as a1,
> lineitem_sf500.l_quantity as a2,
> CAST(lineitem_sf500.l_quantity * lineitem_sf500.l_extendedprice *
> lineitem_sf500.l_discount as double) as a3,
> CAST(lineitem_sf500.l_quantity * lineitem_sf500.l_extendedprice *
> lineitem_sf500.l_tax as double) as a4
> from lineitem_sf500
> ) alias
> group by alias.a0
> ) alias;
> {code}
> The above query was run with ~376GB of data / ~3billion records in the source.
> It takes ~10 minutes with hive.map.aggr=false.
> With map side aggregation set to true, the map tasks don't complete even
> after an hour.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
