asfimport
commented
4 years ago Adrien Grand (@jpountz) (migrated from JIRA)
I'm getting a small but consistently reproducible speedup for boolean/term queries, so I believe it is not noise. Here is the output of one run on wikibigall for instance:
TaskQPS baseline StdDev QPS patch StdDev Pct diff
SpanNear 1.37 (20.9%) 1.36 (20.7%) -0.5% ( -34% - 51%)
PKLookup 195.89 (3.5%) 195.36 (3.8%) -0.3% ( -7% - 7%)
Prefix3 59.17 (6.7%) 59.13 (6.6%) -0.1% ( -12% - 14%)
HighTermDayOfYearSort 46.46 (5.8%) 46.44 (6.1%) -0.1% ( -11% - 12%)
HighTermMonthSort 65.80 (12.4%) 65.80 (12.3%) -0.0% ( -22% - 28%)
Fuzzy1 160.52 (12.7%) 160.58 (13.2%) 0.0% ( -22% - 29%)
IntervalsOrdered 10.76 (3.3%) 10.76 (3.2%) 0.1% ( -6% - 6%)
Wildcard 101.27 (3.7%) 101.33 (4.2%) 0.1% ( -7% - 8%)
SloppyPhrase 6.32 (7.2%) 6.33 (7.2%) 0.1% ( -13% - 15%)
Fuzzy2 80.13 (8.3%) 80.42 (9.2%) 0.4% ( -15% - 19%)
AndHighOrMedMed 37.69 (2.2%) 37.90 (1.9%) 0.6% ( -3% - 4%)
Phrase 10.95 (2.5%) 11.03 (2.4%) 0.8% ( -4% - 5%)
AndMedOrHighHigh 28.77 (2.7%) 29.14 (3.1%) 1.3% ( -4% - 7%)
IntNRQ 92.24 (3.1%) 94.13 (3.4%) 2.1% ( -4% - 8%)
AndHighMed 51.84 (3.0%) 52.92 (3.3%) 2.1% ( -4% - 8%)
Term 1375.94 (2.3%) 1405.17 (3.1%) 2.1% ( -3% - 7%)
OrHighMed 71.51 (2.3%) 73.32 (2.9%) 2.5% ( -2% - 7%)
OrHighHigh 86.33 (2.2%) 89.26 (3.0%) 3.4% ( -1% - 8%)
AndHighHigh 36.11 (2.9%) 37.34 (3.9%) 3.4% ( -3% - 10%)
We changed the way BM25 scores are computed in #9045 in order to guarantee monotonicity of scores, but this translated to a small decrease of throughput, see annotation CC (October 2017) on Mike's nightly benchmarks. Even though the total number of score computations has decreased since we introduced block-max WAND, its relative cost is not negligible since we not only compute scores on collected documents, but also when decoding skip lists in order to compute the maximum score per block, or group of blocks.
Migrated from LUCENE-9071 by Adrien Grand (@jpountz), resolved Dec 09 2019