| commit | d9c62a63dce55a5aaa1b60582fc66a3082109fe1 | [log] [tgz] |
|---|---|---|
| author | Lalit Maganti <lalitm@google.com> | Tue Nov 05 15:43:50 2019 +0000 |
| committer | Lalit Maganti <lalitm@google.com> | Tue Nov 05 15:43:50 2019 +0000 |
| tree | 28c97222e304f35312801d2a8bb6e0ccb7c97573 | |
| parent | 51a6086ce618d4f2c2fed145b939c56f1697eb80 [diff] |
trace_processor: make filtering of non-null columns faster This CL adds a fast-path when filtering non-null columns by skipping the lookup of the bitvector inside SparseVector when we know that every entry of the column is non-null. As most performance sensitive columns will be non-null, it is important that this is efficient. Before: ---------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------------- BM_TableFilterRootNonNullEqMatchMany/1024 13329 ns 13329 ns 52480 BM_TableFilterRootNonNullEqMatchMany/4096 71338 ns 71338 ns 9873 BM_TableFilterRootNonNullEqMatchMany/32768 456830 ns 456832 ns 1532 BM_TableFilterRootNonNullEqMatchMany/262144 3555187 ns 3555110 ns 194 BM_TableFilterRootNonNullEqMatchMany/2097152 27917977 ns 27918146 ns 25 BM_TableFilterChildNonNullEqMatchMany/1024 18743 ns 18743 ns 37067 BM_TableFilterChildNonNullEqMatchMany/4096 98913 ns 98914 ns 7084 BM_TableFilterChildNonNullEqMatchMany/32768 697157 ns 697161 ns 1031 BM_TableFilterChildNonNullEqMatchMany/262144 5231207 ns 5230762 ns 133 BM_TableFilterChildNonNullEqMatchMany/2097152 41994637 ns 41994799 ns 17 BM_TableFilterChildNonNullEqMatchManyInParent/1024 22546 ns 22546 ns 30982 BM_TableFilterChildNonNullEqMatchManyInParent/4096 111175 ns 111176 ns 6320 BM_TableFilterChildNonNullEqMatchManyInParent/32768 790285 ns 790285 ns 884 BM_TableFilterChildNonNullEqMatchManyInParent/262144 6088922 ns 6088676 ns 114 BM_TableFilterChildNonNullEqMatchManyInParent/2097152 55033399 ns 55032881 ns 10 After: ---------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations ---------------------------------------------------------------------------------------------- BM_TableFilterRootNonNullEqMatchMany/1024 9661 ns 9661 ns 72434 BM_TableFilterRootNonNullEqMatchMany/4096 55301 ns 55301 ns 12180 BM_TableFilterRootNonNullEqMatchMany/32768 336006 ns 336009 ns 2098 BM_TableFilterRootNonNullEqMatchMany/262144 2502308 ns 2502299 ns 277 BM_TableFilterRootNonNullEqMatchMany/2097152 19694850 ns 19694393 ns 35 BM_TableFilterChildNonNullEqMatchMany/1024 15019 ns 15019 ns 46872 BM_TableFilterChildNonNullEqMatchMany/4096 83278 ns 83278 ns 8416 BM_TableFilterChildNonNullEqMatchMany/32768 549981 ns 549987 ns 1281 BM_TableFilterChildNonNullEqMatchMany/262144 4190223 ns 4190245 ns 166 BM_TableFilterChildNonNullEqMatchMany/2097152 33867350 ns 33867514 ns 20 BM_TableFilterChildNonNullEqMatchManyInParent/1024 19062 ns 19062 ns 36507 BM_TableFilterChildNonNullEqMatchManyInParent/4096 99900 ns 99898 ns 6967 BM_TableFilterChildNonNullEqMatchManyInParent/32768 684101 ns 684103 ns 1028 BM_TableFilterChildNonNullEqMatchManyInParent/262144 5280275 ns 5280263 ns 132 BM_TableFilterChildNonNullEqMatchManyInParent/2097152 42715263 ns 42715753 ns 16 Analysis: Across all the non-null benchmarks, the new code is 20-30% faster with no regression on any of the other codepaths. This is a good speed up given that we are reaching the limits of how optimized filtering can be. Context: go/perfetto-tp-refactor Bug: 135177627 Change-Id: Ic49c582f406f60d6d24599db7b0eda5b0f1bc6eb
Perfetto is an open-source project for performance instrumentation and tracing of Linux/Android/Chrome platforms and user-space apps.
See www.perfetto.dev for docs.
You can reach us on our Discord channel. If you prefer using IRC we have an experimental Discord <> IRC bridge synced with #perfetto-dev on Freenode.