Paul Rogers created DRILL-5125:
----------------------------------
Summary: Provide option to use generic code for sv remover
Key: DRILL-5125
URL: https://issues.apache.org/jira/browse/DRILL-5125
Project: Apache Drill
Issue Type: Improvement
Affects Versions: 1.8.0
Reporter: Paul Rogers
Assignee: Paul Rogers
Priority: Minor
Consider a non-typical Drill query: one with 6000 rows but 243 fields. Consider
this query:
{code}
select * from (select *, row_number() over(order by somedate) as rn from
dfs.`/some/path/data.json`) where rn=10
{code}
This produces a query with the following structure:
{code}
TBD
{code}
Instrumenting, the code to measure compile time, two “long poles” stood out:
{code}
Compile Time for org.apache.drill.exec.test.generated.CopierGen3: 500
Compile Time for org.apache.drill.exec.test.generated.CopierGen8: 1659
{code}
Much of the initial run time of 5578 ms is taken up in compiling two classes
(2159 ms).
The classes themselves are very simple: create member variables for 486 vectors
(2 x column count), and call a method on each to do the copy. The only
type-specific work is the member variable and call to the (non-virtual)
CopyFrom or CopyFromSafe methods. The generated class can easily be replaced by
a “generic” class and virtual functions in the vector classes to choose the
correct copy method.
Clearly, avoiding code gen means avoiding the compile times with a first-run
savings. Here are the last 8 runs (out of 10), with code cached turned off
(forcing a compile on each query run), with and without the generic versions:
* Original (no code cache): 1832 ms / run
* Generic (no code cache): 1317 ms / run
This demonstrates the expected outcome: avoiding compilation of generated code
saves ~500 ms per run (or 28%). (Note: the numbers above were obtained on a
version of the code that already had various optimizations described in other
JIRA entries.)
The reason, for generating code is that one would expect that 243 in-line
statements (an unwound loop) to be faster than a loop with 243 iterations. In
addition, the generic version uses an array in place of ~500 variables, and a
virtual function call rather than in-line, type-specific calls. One would
expect the unrolled loop to be faster.
Repeat the exercise, this time with the code cache turned on so that no compile
cost is payed for either code path (because the test excludes the first two
runs in which the generated code is compiled.)
* Original: 1302 ms / run
* Generic version: 1040 ms / run
Contrary to expectations, the loop is faster than the in-line statements. In
this instance, the array/loop/virtual function version is ~260 ms faster (20%).
The test shows that the code can be simplified, a costly costly code-gen and
compile step can be skipped, and this query will go faster. Plus, since the
change removes generated classes from the code cache, there is more room for
the remaining classes, which may improve the hit rate.
This ticket offers the performance improvement as an option, described in
comments.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
