easyice commented on PR #12782: URL: https://github.com/apache/lucene/pull/12782#issuecomment-1806834098
@jpountz You are right, recomputing the length is faster than table lookup, here is the benchmark when reading the ints, each value will takes 4 bytes: ``` GroupVInt.readGroupVInt 16 thrpt 5 11.822 ± 0.187 ops/us GroupVInt.readGroupVInt 32 thrpt 5 7.558 ± 0.209 ops/us GroupVInt.readGroupVInt 64 thrpt 5 3.556 ± 0.344 ops/us GroupVInt.readGroupVInt 128 thrpt 5 1.786 ± 0.145 ops/us GroupVInt.readGroupVInt 248 thrpt 5 0.972 ± 0.025 ops/us GroupVInt.readGroupVIntWithoutTable 16 thrpt 5 19.787 ± 2.848 ops/us GroupVInt.readGroupVIntWithoutTable 32 thrpt 5 10.162 ± 1.491 ops/us GroupVInt.readGroupVIntWithoutTable 64 thrpt 5 5.141 ± 0.121 ops/us GroupVInt.readGroupVIntWithoutTable 128 thrpt 5 2.247 ± 0.017 ops/us GroupVInt.readGroupVIntWithoutTable 248 thrpt 5 1.183 ± 0.014 ops/us GroupVInt.readVInt 16 thrpt 5 12.679 ± 0.405 ops/us GroupVInt.readVInt 32 thrpt 5 6.519 ± 0.247 ops/us GroupVInt.readVInt 64 thrpt 5 3.218 ± 0.804 ops/us GroupVInt.readVInt 128 thrpt 5 1.762 ± 0.096 ops/us GroupVInt.readVInt 248 thrpt 5 0.887 ± 0.035 ops/us ``` But i found that the group-varint encoding is faster than vint only when the value takes up 4 bytes, here is the benchmark for reading 64 int values. `numBytesOfInt` is the bytes for an int value that will be taken : ``` Benchmark (numBytesOfInt) (size) Mode Cnt Score Error Units GroupVInt.readGroupVIntWithoutTable 1 64 thrpt 5 5.099 ± 0.147 ops/us GroupVInt.readGroupVIntWithoutTable 2 64 thrpt 5 4.982 ± 0.632 ops/us GroupVInt.readGroupVIntWithoutTable 3 64 thrpt 5 5.194 ± 0.163 ops/us GroupVInt.readGroupVIntWithoutTable 4 64 thrpt 5 4.923 ± 0.092 ops/us GroupVInt.readVInt 1 64 thrpt 5 8.433 ± 0.287 ops/us GroupVInt.readVInt 2 64 thrpt 5 6.309 ± 0.155 ops/us GroupVInt.readVInt 3 64 thrpt 5 5.196 ± 0.213 ops/us GroupVInt.readVInt 4 64 thrpt 5 3.300 ± 0.302 ops/us ``` The decoding for group-varint takes up constant time. vint decoding is faster when the values take up fewer bytes. so the actual payoff depends on factors like `maxDoc`. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
