perf: implement bitpack encoding for LID and MID blocks#328
Conversation
3accfd6 to
53ef5eb
Compare
|
@seqbenchbot up main bulk |
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
|
@seqbenchbot down dc2d4d40 |
|
Nice, @cheb0 The benchmark with identificator Show summary
Have a great time! |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
bf8f3d8 to
985f11f
Compare
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
3a567c4 to
f3827f1
Compare
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
|
@seqbenchbot up main bulk |
|
@seqbenchbot down d9e8e68e |
|
Nice, @cheb0 The benchmark with identificator Show summary
Have a great time! |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
🔴 Performance DegradationSome benchmarks have degraded compared to the previous run. Show table
|
Description
Replaces varint encoding with faster delta bitpacking. Both LID and MID blocks now use bitpack. Currently, intcomp library is used. The lib doesn't utilize SIMD, so we might update to something else in future.
Measurements
compression
bitpack compresses a lot better: for varints zstd compresses with ratio ~1.7-2.0 while it only compresses delta bitpacked data with ratio ~1.3. Therefore we potentially can disable zstd on benchmarks with a slight dataset size overhead.
dataset size
Overall, we reach approximately same dataset size. For some envs there is a small benefit of around -3% of total dataset.
Search latency (stg fractions repacked with bitpack)
Fixes #312