[ENHANCEMENT]: Add XXHash
sleeepyjack opened this issue · comments
Is your feature request related to a problem? Please describe.
The current default hasher murmur3_32 is computational expensive and might lead to suboptimal performance in libcudf's hash join for when the build table is small enough so that it fits in L1 or L2$ (see rapidsai/cudf#10587).
Describe the solution you'd like
Libcudf has an open feature request (see rapidsai/cudf#12829), which proposes to use a faster hash function, namely XXHash. I propose adding XXHash as a second hasher to cuco and subsequently compare performance against murmur3_32.
Describe alternatives you've considered
No response
Additional context
No response
Awesome @sleeepyjack! Feel free to tag me for review if you'd like. I'd be very interested in seeing what libcudf can improve here.