feat: faster bitpacking filter for selectivities from 5% to 80% #2068
+107
−6
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Benchmarks: random_access |
Benchmarks: datafusionTable of Results
|
Benchmarks: TPC-HTable of Results
|
Benchmarks: ClickbenchTable of Results
|
|
It's worth running this on our AVX512 machine too, to see if the switch point depends on SIMD width. M3's only have 128 bits IIRC. |
|
On a c2-standard-4 (Cascade Lake), the switch points are slightly different. It seems i8: 0.02, i16: 0.03, i32: 0.075, i64: 0.09. This PR uses: i8: 0.02, i16: 0.03, i32: 0.04, i64: 0.04. 512 / 128 = 4. These tests use 10,000 element arrays, so 0.04 is around 400 elements whereas 0.075 and 0.09 are around 750 and 900. I'm not sure there's a robust way to pick this threshold without benchmarking on the target machine. I'd be happy to push the 4 and 8 byte types up to 0.075 and 0.09. On an Apple M3 this is 20-35% slower but we're talking about 2.0 us vs 1.5 us. CPU |
…5% and 80% This new benchmark demonstrates that the switchpoint is in [0.02, 0.04]. 8-bit elements switch around 0.02, but 32- and 64-bit elements switch around 0.04. [Google Sheet with one run on my Apple M3 Max](https://docs.google.com/spreadsheets/d/1T4JeSLnpFegA_pRS70iNu4ve9YMjEu-j1vRL7spazoA/edit?gid=624487667#gid=624487667).
danking
force-pushed
the
dk/bitpacking-filter-selection-threshold
branch
from
02cab19 to
369f8cf
Compare
|
Okay, I went with the Cascade Lake threshold as those are best for our benchmarks. I wish I had a more principled way to write them down or some way to tune to the current CPU. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
On an Apple M3 Max, the new benchmark in this PR demonstrates that the switchpoint is in [0.02, 0.04]. 8-bit, 16-bit, 32-bit, and 64-bit arrays switch at 0.02, 0.03, 0.04, and 0.04, respectively. I later ran the benchmarks on a Cascade Lake cloud machine and found that the 8-bit, 16-bit, 32-bit, and 64-bit arrays switch at 0.03, 0.03, 0.075 and 0.09, respectively. In this PR, I use the Cascade Lake values, but I don't have a great answer for how to pick these. Regardless, it is clear that 0.8 is not the correct choice.
Google Sheet with one run on my Apple M3 Max. A sheet with a run on a c2-standard-4 "Cascade Lake"