You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I would like to know if switching our default hasher from MurmurHash3_32 to XXHash32 has an impact on performance. This was previously proposed in #12829 but deserves an issue of its own to be acted upon.
If the register usage is lower for XXHash32 than MurmurHash3_32, it might benefit some of cudf's most complex kernels. Mixed joins, for instance, combine hashing and AST execution. Similar ideas have been mentioned before in issues like #10587. I'm not confident that this would make a positive difference (I expect no significant effect, based on our past explorations of mixed joins), but it would be worth checking.
I would like to know if switching our default hasher from MurmurHash3_32 to XXHash32 has an impact on performance. This was previously proposed in #12829 but deserves an issue of its own to be acted upon.
If the register usage is lower for XXHash32 than MurmurHash3_32, it might benefit some of cudf's most complex kernels. Mixed joins, for instance, combine hashing and AST execution. Similar ideas have been mentioned before in issues like #10587. I'm not confident that this would make a positive difference (I expect no significant effect, based on our past explorations of mixed joins), but it would be worth checking.
Originally posted by @bdice in #12829
The text was updated successfully, but these errors were encountered: