Closed fknorr closed 6 months ago
Check-perf-impact results: (f7ef1830fe7f830eee417b0b5a37c2ef)
:warning: Significant slowdown (>1.25x) in some microbenchmark results: building command graphs in a dedicated scheduler thread for N nodes - 1 > immediate submission to a scheduler thread / contracting tree topology
:rocket: Significant speedup (<0.80x) in some microbenchmark results: building command graphs in a dedicated scheduler thread for N nodes - 4 > immediate submission to a scheduler thread / contracting tree topology
Relative execution time per category: (mean of relative medians)
Files with Coverage Reduction | New Missed Lines | % | ||
---|---|---|---|---|
src/grid.cc | 1 | 97.29% | ||
<!-- | Total: | 1 | --> |
Totals | |
---|---|
Change from base Build 7875546105: | -0.03% |
Covered Lines: | 4999 |
Relevant Lines: | 5194 |
Note: This PR is based on #248.
In #234, we set the global log level to
trace
in all tests in order to print a full log in every failing test. Since we did not perform a benchmark run, we did not notice its impact on our benchmarking suite (see the benchmark report in #248): Task-graph generation emits trace-logs whenever testing for horizon generation, which is triggered for every command group submission. These logs are being written to the test log capture buffer which is very expensive compared to the actual task generation.This PR explicitly shuts off all logging for the duration of any benchmark, re-activating the behavior that was implicitly true before #248 (because benchmarks were run with the default
LOG_LEVEL
).I have re-run the benchmarks to commit a new baseline.