[
https://issues.apache.org/jira/browse/IMPALA-3360?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Thomas Tauber-Marshall resolved IMPALA-3360.
--------------------------------------------
Resolution: Fixed
Fix Version/s: Impala 2.11.0
commit 79dc220bd75eb5dc333aeeff3f65fc5dbfe3a6e8
Author: Thomas Tauber-Marshall <tmarsh...@cloudera.com>
Date: Wed Sep 6 12:29:38 2017 -0700
IMPALA-3360: Codegen inserting into runtime filters
This patch codegens PhjBuilder::InsertRuntimeFilters() and
FilterContext::Insert().
This allows us to unroll the loop over all the filters in
PhjBuilder::ProcessBuildBatch(), eliminate the branch on type that
happens in RawValue::GetHashValue(), and eliminate the AVX check
that happens in BloomFilter::Insert().
Testing:
- Ran existing runtime filter tests.
- Ran perf tests locally (all avg. over three runs):
- Four way self join on tpch_parquet.lineitem. Should be a good case
for this as there's several large hash join build sides that will
benefit from the codegen. Total query running time improved ~7%
(from 16.07s to 14.91s).
- Single join of tpch_parquet.lineitem against a selectively
filtered tpch_parquet.lineitem. Should be a bad case for this
patch, as the build side of the join is very small. Total query
running time regressed by about ~2% (from 0.73s to 0.75s) due to
an increase in codegen time (from 295ms to 309ms for the fragment
containing the hash join).
Change-Id: I79cf23ad92dadaab996a50a2ca07ef9ebe8639bb
Reviewed-on: http://gerrit.cloudera.org:8080/8029
Reviewed-by: Thomas Tauber-Marshall <tmarsh...@cloudera.com>
Tested-by: Impala Public Jenkins
> Unroll loops / replace types in filter logic in PHJ::ProcessBuildBatch()
> ------------------------------------------------------------------------
>
> Key: IMPALA-3360
> URL: https://issues.apache.org/jira/browse/IMPALA-3360
> Project: IMPALA
> Issue Type: Bug
> Components: Backend
> Affects Versions: Impala 2.6.0
> Reporter: Henry Robinson
> Assignee: Thomas Tauber-Marshall
> Priority: Minor
> Labels: codegen
> Fix For: Impala 2.11.0
>
>
> This code can be optimized for codegen:
> {code}
> for (const FilterContext& ctx: filters_) {
> if (ctx.local_bloom_filter == NULL) continue;
> void* e = ctx.expr->GetValue(build_row);
> uint32_t filter_hash = RawValue::GetHashValue(e,
> ctx.expr->root()->type(),
> RuntimeFilterBank::DefaultHashSeed());
> ctx.local_bloom_filter->Insert(filter_hash);
> }
> {code}
> Note also that we know ahead of time whether {{ctx.local_bloom_filter}} will
> be {{NULL}}, so we could prepare a list of non-null filters and only iterate
> over that.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
