Repository navigation
Add Delta decode, probe, filter and take benchmarks - #11429
Conversation
delta_decode measures a u32/u16/u64 Delta array over bit-packed deltas: decoding it with the deltas packed and unpacked, reading rows through one-off and repeated probes, and filtering and taking from it compared with decoding first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018SWbErHBxmZXBVmmLcAbLD Signed-off-by: Robert Kruszewski <github@robertk.io>
Merging this PR will improve performance by 15.64%3 warnings
Performance Changes
Tip Curious why performance improved? Comment Comparing Footnotes
|
joseph-isaacs
left a comment
There was a problem hiding this comment.
48 benchmarks is too much
Only decode_bitpacked_deltas runs on each walltime CPU-feature leg, since the build picks its transpose and decode kernels. The probe, filter and take benchmarks run once, in simulation. decode_unpacked_deltas becomes decode_primitive_deltas, which decodes a Delta array over primitive deltas without unpacking them inside the measurement, and the decode-then-filter and decode-then-take baselines are dropped. That leaves 15 benchmarks instead of 48. The six simulation benchmarks go into scripts/codspeed_bench_counts.json, which the shard planner requires to list every benchmark target. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018SWbErHBxmZXBVmmLcAbLD Signed-off-by: Robert Kruszewski <github@robertk.io>
Simulation runs roughly 13x slower than native, so the probe benchmarks over a 512 KiB array took ~150 ms. Use 32 KiB arrays for the simulation-only benchmarks, read 8 rows from a 2048-row array in the probe benchmarks, and keep 512 KiB only for the walltime decode. Signed-off-by: Robert Kruszewski <github@robertk.io> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018SWbErHBxmZXBVmmLcAbLD
This PR adds benchmarks for the Delta encoding. They are split out of #11428 so that its changes can be measured against
develop, and #11428 stacks on this PR.