pcb-rnd knowledge pool
Query() performance test 2020
| query_speed by Tibor 'Igor2' Palinkas on 2020-05-16 | Tags: insight, query, performance, speed, expression, search |
Abstract: n/a
Test environment
The test script was running on an intel i5-3470 @ 3.2Ghz, circa 2013. The test board has a grid of 10000 identical line objects. Results are the 'user' line of bash 'time' output and represent 1000 pairs of select/unselect operation matching all 10k objects on the board. The executable was pcb-rnd r31411 compiled with -O3.
Test 1: Iteration speed
The first test compares query() setup cost and iteration speed to conventional Select(All) and Unselect(All). This gives a picture on the overhead of the query() infrastructure.
| test case | total run time [s] |
|---|---|
| Select()/Unselect() | 1.967 |
| query, @ | 0.940 |
| query, @.type == LINE | 1.339 |
| query, (@.type == LINE) && (@.thickness > 0.1mil) && (@.thickness < 1000mil) && (@.layer.type == COPPER) | 4.753 |
Test 2: String handling
The second test measures string comparison and regex matching speed against against plain numeric queries above.
| test case | total run time [s] |
|---|---|
| query, (@.layer.type == COPPER) | 1.616 |
| query, (@.layer.name == "top-sig") | 1.542 |
| query, (@.layer.name ~ "top-sig") | 1.858 |
Conclusion
The fixed overhead of the query() system for object listing is lower than the overhead of the rtree search in Select(). Expression evaluation cost is measurable, although not drastic for the usual short searches. The cost of regex matching is only sligtly higher than the cost of plain string comparison.