[
https://issues.apache.org/jira/browse/CALCITE-4494?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Igor Lozynskyi updated CALCITE-4494:
------------------------------------
Description:
*Problem*
Currently, the planning process shows a performance degradation when comparing
to version 1.25. Worse palling time seems to affect most queries, but it is
especially clear for queries with many Rel nodes (especially with multiple
joins).
In a downstream project, we have a stress test that checks the planning time.
In some cases, the planning time is increased by x4 (for a query with 28 joins).
The main contributing factor (but not the only one) for the slow-down could be
traced to [https://github.com/apache/calcite/pull/2222/files].
*Potential Solution*
As it was mentioned by the reviewers, we may improve the current situation with
some tiny changes:
* Introduce a method to check that a Rel node belongs to the RelSubset instead
of getting all Rel nodes (the current code may take up to 60% of the planning
time).
* Improve the null check in RelMdPredicates by building an error message in
RelMdPredicates.ExprsItr only when it is required (may additionally take 10% of
the planning time due to SortedMap.toString() being expensive when frequently
called).
With these 2 changes, I was able to regain most of the lost planning
performance.
The following flame graph clearly shows that the call to RelSubset.getRelList()
from VolcanoRuleCall.onMatch() is expensive (28 join query):
!CalcitePerf_Planning_RelList_consumes_a_lot.png|width=865,height=365!
After the proposed improvements, the flame graph shows the following (28 join
query):
!CalcitePerf_Planning_after_improvements.png|width=561,height=472!
It is clear that the HintStrategyTable.isRuleExcluded() call is expensive, but
the overall picture is much better.
Also, in my environment, the TPC-H Q7 test takes ~20% less time (39.6 sec vs
32.9 sec) after the proposed improvements. Here, the flame graph also shows
that ordinary queries are also affected by the redundant RelSubset.getRelList()
calls:
!CalcitePerf_Planning_TPCH_Q7_RelList_consumes_a_lot.png|width=856,height=573!
was:
*Problem*
Currently, the planning process shows a performance degradation when comparing
to version 1.25. Worse palling time seems to affect most queries, but it is
especially clear for queries with many Rel nodes (especially with multiple
joins).
In a downstream project, we have a stress test that checks the planning time.
In some cases, the planning time is increased by x4 (for a query with 28 joins).
The main contributing factor (but not the only one) for the slow-down could be
traced to
[https://github.com/apache/calcite/pull/2222/files|[https://github.com/apache/calcite/pull/2222/files].]
*Potential Solution*
As it was mentioned by the reviewers, we may improve the current situation with
some tiny changes:
* Introduce a method to check that a Rel node belongs to the RelSubset instead
of getting all Rel nodes (the current code may take up to 60% of the planning
time).
* Improve the null check in RelMdPredicates by building an error message in
RelMdPredicates.ExprsItr only when it is required (may additionally take 10% of
the planning time due to SortedMap.toString() being expensive when frequently
called).
With these 2 changes, I was able to regain most of the lost planning
performance.
The following flame graph clearly shows that the call to RelSubset.getRelList()
from VolcanoRuleCall.onMatch() is expensive:
!CalcitePerf_Planning_RelList_consumes_a_lot.png!
After the proposed improvements, the flame graph shows the following:
!CalcitePerf_Planning_after_improvements.png!
It is clear that the HintStrategyTable.isRuleExcluded() call is expensive, but
the overall picture is much better.
Also, in my environment, the TPC-H Q7 test takes ~20% less time (39.6 sec vs
32.9 sec) after the proposed improvements. Here, the flame graph also shows
that ordinary queries are also affected by the redundant RelSubset.getRelList()
calls:
!CalcitePerf_Planning_TPCH_Q7_RelList_consumes_a_lot.png!
> Improve planning performance with RelSubset check for Rel presence
> -------------------------------------------------------------------
>
> Key: CALCITE-4494
> URL: https://issues.apache.org/jira/browse/CALCITE-4494
> Project: Calcite
> Issue Type: Improvement
> Components: core
> Affects Versions: 1.26.0
> Environment: All environments
> Reporter: Igor Lozynskyi
> Priority: Major
> Labels: performance
> Fix For: 1.27.0
>
> Attachments: CalcitePerf_Planning_RelList_consumes_a_lot.png,
> CalcitePerf_Planning_TPCH_Q7_RelList_consumes_a_lot.png,
> CalcitePerf_Planning_after_improvements.png
>
>
> *Problem*
> Currently, the planning process shows a performance degradation when
> comparing to version 1.25. Worse palling time seems to affect most queries,
> but it is especially clear for queries with many Rel nodes (especially with
> multiple joins).
> In a downstream project, we have a stress test that checks the planning time.
> In some cases, the planning time is increased by x4 (for a query with 28
> joins).
> The main contributing factor (but not the only one) for the slow-down could
> be traced to [https://github.com/apache/calcite/pull/2222/files].
> *Potential Solution*
> As it was mentioned by the reviewers, we may improve the current situation
> with some tiny changes:
> * Introduce a method to check that a Rel node belongs to the RelSubset
> instead of getting all Rel nodes (the current code may take up to 60% of the
> planning time).
> * Improve the null check in RelMdPredicates by building an error message in
> RelMdPredicates.ExprsItr only when it is required (may additionally take 10%
> of the planning time due to SortedMap.toString() being expensive when
> frequently called).
> With these 2 changes, I was able to regain most of the lost planning
> performance.
> The following flame graph clearly shows that the call to
> RelSubset.getRelList() from VolcanoRuleCall.onMatch() is expensive (28 join
> query):
> !CalcitePerf_Planning_RelList_consumes_a_lot.png|width=865,height=365!
> After the proposed improvements, the flame graph shows the following (28 join
> query):
> !CalcitePerf_Planning_after_improvements.png|width=561,height=472!
> It is clear that the HintStrategyTable.isRuleExcluded() call is expensive,
> but the overall picture is much better.
> Also, in my environment, the TPC-H Q7 test takes ~20% less time (39.6 sec vs
> 32.9 sec) after the proposed improvements. Here, the flame graph also shows
> that ordinary queries are also affected by the redundant
> RelSubset.getRelList() calls:
> !CalcitePerf_Planning_TPCH_Q7_RelList_consumes_a_lot.png|width=856,height=573!
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
