Tim Armstrong has posted comments on this change. Change subject: Add FNV, Zobrist, and SIMD hash functions to the int hash benchmark. ......................................................................
Patch Set 1: (3 comments) Thanks, this is really interesting. http://gerrit.cloudera.org:8080/#/c/4313/1/be/src/benchmarks/int-hash-benchmark.cc File be/src/benchmarks/int-hash-benchmark.cc: Line 41: // FNV 25.7 26 26.4 1X 1X 1X Do we have a sense for the relative quality of the different hash functions? It's hard to interpret the merits of them without understanding the trade-offs. Line 45: // Jenkins1 51.8 54 54 2.02X 2.08X 2.05X Any reason Multiple<Jenkins..> was left out? It looks competitive based on these numbers. Line 271: unique_ptr<uint32_t[]> ud(new uint32_t[DATA_SIZE]); It'd be interesting to see how the vectorised hashing performs if the input data is non-contiguous, e.g. ExprValuesCache.expr_values_array_ might have multiple values to hash per input row. (I guess we could also just transpose that array if we wanted to vectorise it). -- To view, visit http://gerrit.cloudera.org:8080/4313 To unsubscribe, visit http://gerrit.cloudera.org:8080/settings Gerrit-MessageType: comment Gerrit-Change-Id: I11d48f8816d5b129858a1f773015e51049dd1d61 Gerrit-PatchSet: 1 Gerrit-Project: Impala-ASF Gerrit-Branch: master Gerrit-Owner: Jim Apple <[email protected]> Gerrit-Reviewer: Tim Armstrong <[email protected]> Gerrit-HasComments: Yes
