Project : MariaDB ColumnStore MCOL-4995
This project goal is to implement basic vectorized filtering for Mariadb Column Store Engine.
During my participation in the Mariadb as a GSoC2022 student.I have made serveral tasks.
-
Implement vectorized filtering for MCS. This is the main task.I use serveral ways to implement it.
I need to unify arm neon and SSE to provide the same interface for upper layer.Due to that ,it has a little performance loss.If we want to unify the behavior of simdProcessor, we have to make concessions.It is a trade-off.Then I implement vectorized filtering for MCS.
We get some excellent result from micro benchmark.
Before my work. ----------------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ----------------------------------------------------------------------------------------------------- FilterBenchFixture/BM_ColumnScan1ByteVectorizedCode 82923 ns 82925 ns 8351 FilterBenchFixture/BM_ColumnScan1Byte1FilterVectorizedCode 117027 ns 117044 ns 5949 FilterBenchFixture/BM_ColumnScan2ByteTemplatedCode 41096 ns 41098 ns 17019 FilterBenchFixture/BM_ColumnScan2Byte1FilterTemplatedCode 58552 ns 58570 ns 11999 FilterBenchFixture/BM_ColumnScan2Byte1FilterVectorizedCode 58474 ns 58490 ns 11934 FilterBenchFixture/BM_ColumnScan4ByteTemplatedCode 21984 ns 21986 ns 31855 FilterBenchFixture/BM_ColumnScan4ByteVectorizedCode 29395 ns 29413 ns 23626 FilterBenchFixture/BM_ColumnScan8ByteTemplatedCode 10895 ns 10896 ns 64333 FilterBenchFixture/BM_ColumnScan8Byte1FilterTemplatedCode 14439 ns 14453 ns 47839 FilterBenchFixture/BM_ColumnScan8ByteVectorizedCode 10841 ns 10842 ns 63192 FilterBenchFixture/BM_ColumnScan8Byte1FilterVectorizedCode 14414 ns 14428 ns 48412 After the vectorized filtering has been implemented. ----------------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ----------------------------------------------------------------------------------------------------- FilterBenchFixture/BM_ColumnScan1ByteVectorizedCode 12492 ns 12495 ns 56003 FilterBenchFixture/BM_ColumnScan1Byte1FilterVectorizedCode 16263 ns 16279 ns 42995 FilterBenchFixture/BM_ColumnScan2ByteTemplatedCode 10872 ns 10875 ns 64348 FilterBenchFixture/BM_ColumnScan2Byte1FilterTemplatedCode 13261 ns 13272 ns 52742 FilterBenchFixture/BM_ColumnScan2Byte1FilterVectorizedCode 13269 ns 13277 ns 52677 FilterBenchFixture/BM_ColumnScan4ByteTemplatedCode 8198 ns 8200 ns 85331 FilterBenchFixture/BM_ColumnScan4ByteVectorizedCode 9381 ns 9392 ns 74545 FilterBenchFixture/BM_ColumnScan8ByteTemplatedCode 6881 ns 6885 ns 101643 FilterBenchFixture/BM_ColumnScan8Byte1FilterTemplatedCode 8574 ns 8595 ns 81450 FilterBenchFixture/BM_ColumnScan8ByteVectorizedCode 6860 ns 6863 ns 101982 FilterBenchFixture/BM_ColumnScan8Byte1FilterVectorizedCode 8588 ns 8603 ns 81351
As we can see,the performance has been improved a lot.And for a such query " select * from tablename where colname='number' " at dataset cardinality is 1 000 000 000.Its performance is improved by approximately 40%.
The related Prs are the follows.
MCOL-4995 Research/implement basic vectorized filtering for ARM platforms
Beyond those,I also use a different way that is suitable for arm neon. And I get more excellent result from that.It is improved about 2x than before according to micro benchmark.
-
FilterBenchFixture/BM_ColumnScan1ByteTemplatedCode 129939 ns 129946 ns 5390 FilterBenchFixture/BM_ColumnScan1Byte1FilterTemplatedCode 131282 ns 131301 ns 5328 FilterBenchFixture/BM_ColumnScan1ByteVectorizedCode 8172 ns 8173 ns 85692 FilterBenchFixture/BM_ColumnScan1Byte1FilterVectorizedCode 8975 ns 8992 ns 77879 FilterBenchFixture/BM_ColumnScan2ByteTemplatedCode 5041 ns 5049 ns 138703 FilterBenchFixture/BM_ColumnScan2Byte1FilterTemplatedCode 5775 ns 5791 ns 120794 FilterBenchFixture/BM_ColumnScan2Byte1FilterVectorizedCode 5770 ns 5783 ns 121063 FilterBenchFixture/BM_ColumnScan4ByteTemplatedCode 4084 ns 4086 ns 171615 FilterBenchFixture/BM_ColumnScan4ByteVectorizedCode 4434 ns 4448 ns 157333 FilterBenchFixture/BM_ColumnScan8ByteTemplatedCode 2779 ns 2781 ns 252110 FilterBenchFixture/BM_ColumnScan8Byte1FilterTemplatedCode 3643 ns 3656 ns 191413 FilterBenchFixture/BM_ColumnScan8ByteVectorizedCode 2781 ns 2782 ns 251334 FilterBenchFixture/BM_ColumnScan8Byte1FilterVectorizedCode 3640 ns 3656 ns 191375
And the result from a query in general is 2x speed up(For some reasons,this result is produced by my mentor).And the data cardinality is 1000000000.
MariaDB [test]> select i from cs1 where i = 5000 or i = 7000;
Empty set (0.338 sec)
MariaDB [test]> select i from cs1 where i = 5000 or i = 7000;
Empty set (0.339 sec)
MariaDB [test]> select i from cs1 where i = 5000 or i = 7000;
Empty set (29.356 sec)
MariaDB [test]> select i from cs1 where i = 5000 or i = 7000;
Empty set (0.724 sec)
MariaDB [test]> select i from cs1 where i = 5000 or i = 7000;
Empty set (0.723 sec)The first two runs are cached with vectorization, the next one that is 29 secs is cold run(slow disks), the last two are with vectorization disabled.
The related code is here.
And I also use SVE to implement vectorized filtering.But the performance is worse than arm neon.
- optimize string comparasions for primproc.
For Eq comparasions in p_Dictionary function of pp,the way uses hashmap is faster than for loop, but when the filtercount is less than a magic number N, the for loop method is faster.And I have simply found that N is 6,but it is just an estimate.
Accroding to this micro benchmark.The time complexity of for loop is O(n).We can find that when filtercount is 3,the time for loop method is half of eqFilter. So when filtercount is less than 6,the way uses for loop benefits a lop to performance.
The related Pr is here
-
Bug fixes
- Try MCOL-4783( Propagate LIMIT into TupleBPS to reduce the number of simultaneous Primitive Jobs issued by TBPS)
I want to express my thanks to drrtuy (Roman Nozdrin) for mentoring and helping me.