[
https://issues.apache.org/jira/browse/DRILL-6211?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kunal Khatua updated DRILL-6211:
--------------------------------
Labels: performance (was: )
> Optimizations for SelectionVectorRemover
> -----------------------------------------
>
> Key: DRILL-6211
> URL: https://issues.apache.org/jira/browse/DRILL-6211
> Project: Apache Drill
> Issue Type: Bug
> Components: Execution - Codegen
> Reporter: Kunal Khatua
> Assignee: Karthikeyan Manivannan
> Priority: Major
> Labels: performance
> Fix For: 1.15.0
>
> Attachments: 255d264c-f55e-b343-0bef-49d3e672d93f.sys.drill,
> 255d2664-2418-19e0-00ea-2076a06572a2.sys.drill,
> 255d2682-8481-bed0-fc22-197a75371c04.sys.drill,
> 255d26ae-2c0b-6cd6-ae71-4ad04c992daf.sys.drill,
> 255d2880-48a2-d86b-5410-29ce0cd249ed.sys.drill
>
>
> Currently, when a SelectionVectorRemover receives a record batch from an
> upstream operator (like a Filter), it immediately starts copying over records
> into a new outgoing batch.
> It can be worthwhile if the RecordBatch can be enriched with some additional
> summary statistics about the attached SelectionVector, such as
> # number of records that need to be removed/copied
> # total number of records in the record-batch
> The benefit of this would be that in extreme cases, if *all* the records in a
> batch need to be either truncated or copies, the SelectionVectorRemover can
> simply drop the record-batch or simply forward it to the next downstream
> operator.
> While the extreme cases of simply dropping the batch kind of works (because
> there is no overhead in copying), for cases where the record batch should
> pass through, the overhead remains (and is actually more than 35% of the
> time, if you discount for the streaming agg cost within the tests).
> Here are the statistics of having such an optimization
> ||Selectivity||Query Time||%Time used by SVR||Time||Profile||
> |0%|6.996|0.13%|0.0090948|[^255d264c-f55e-b343-0bef-49d3e672d93f.sys.drill]|
> |10%|7.836|7.97%|0.6245292|[^255d2682-8481-bed0-fc22-197a75371c04.sys.drill]|
> |50%|11.225|25.59%|2.8724775|[^255d2664-2418-19e0-00ea-2076a06572a2.sys.drill]|
> |90%|14.966|33.91%|5.0749706|[^255d26ae-2c0b-6cd6-ae71-4ad04c992daf.sys.drill]|
> |100%|19.003|35.73%|6.7897719|[^255d2880-48a2-d86b-5410-29ce0cd249ed.sys.drill]|
> To summarize, the SVR should avoid creating new batches as much as possible.
> A more generic (non-trivial) optimization should take into account the fact
> that multiple batches emitted can be coalesced, but we don't currently have
> test metrics for that.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
