Tidb: Vectorize hash calculation in hashJoin.

Created on 5 Sep 2019  路  6Comments  路  Source: pingcap/tidb

Feature Request

Is your feature request related to a problem? Please describe:

In hashJoin, we do the hash calculation for a chunk row by row https://github.com/pingcap/tidb/blob/1ff620d/executor/hash_table.go#L143-L153, can try to vectorize it for performance.

Describe the feature you'd like:

  • [x] vectorize hash calculation during building hash table. #12076 by @sduzh
  • [x] vectorize hash calculation during probing. #12669 by @sduzh

Describe alternatives you've considered:

Teachability, Documentation, Adoption, Migration Strategy:

siexecution statuhelp-wanted typenhancement typperformance

Most helpful comment

I can fix it

All 6 comments

I can fix it

@sduzh Cool

Hi @sduzh, Thanks for your awesome contribution.
Do you have time to do another job?
We can also vectorize the way to calculate hash values for the probe rows.

Hi @sduzh, Thanks for your awesome contribution.
Do you have time to do another job?
We can also vectorize the way to calculate hash values for the probe rows.

Yes, no problem. @SunRunAway

@sduzh A friendly ping, how's it going?

@sduzh A friendly ping, how's it going?

I finished the vectorized hash value calculation of probe phase weeks ago but the benchmark scores are decreased.
I will take an investigation of that this weekend and try to resolve it.

Was this page helpful?
0 / 5 - 0 ratings