tp: keep what executor loops read in locals **Why** In a loop which writes through an integer pointer, or writes bytes, the compiler cannot tell the writes from what the loop reads through a reference, a pointer or a struct field, so it reads each of them again after every write. `Filter` compared every row with a value it read through a reference: an equality scan of the counter table ran 23.4G instructions where the interpreter ran 13.9G. Clang's `-Rpass-missed=licm` remarks found the same pattern in the per-row loops of most operators. **Changes** - Loops which run once per row copy their bounds, the values they compare with and the fields they read into locals first. This covers `Filter`, the row layout sort, `ColumnChunk`'s copies and gathers, `Expand`, `DataframeWindows`, the aggregate functions, hash aggregation, `AssertType`, the interval, graph and tree operators and `TopK`. - `Selection::indices()` returns the rows a selection keeps, so a loop tests `prefix()` once rather than for every row. `FlatColumnReader` keeps them rather than the selection. - `ColumnChunk::Gather`'s loop moves out of its lambda into `GatherValues()`, as the lambda's captures were read through the closure. - `GraphTree` fills its node, seed and edge lists through presized vectors rather than calling `push_back()` in its loops. **Benchmarks** Release build, median of 7 runs, two rounds: | Benchmark | Change | |---|---:| | `BM_ExecFilterEqNullable` | -22% | | `BM_SortRowLayoutSorted` | -13% to -15% | | `BM_TreeNumberNodesIdColumn/10000` | -12% | | `BM_ExecFilterEqUint32` | -4% | The equality scan of the counter table runs 14.7G instructions. **Testing** No behaviour change. `perfetto_unittests` and all diff tests pass.
Perfetto is an open-source suite of tools for recording, visualizing and analyzing traces: timelines of what software and the system under it were doing. It is the default tracing system for Android and Chromium. System-wide recording works on Android and Linux. The tracing SDK, the UI and Trace Processor also run on macOS and Windows.
perf, Firefox Profiler, macOS Instruments, pprof, systrace, Fuchsia and more.Open an example Android trace without installing anything, or drag your own trace file onto ui.perfetto.dev.
On Android, with a device connected over adb:
curl -O https://raw.githubusercontent.com/google/perfetto/main/tools/record_android_trace python3 record_android_trace -o trace.pftrace -t 10s sched freq gfx view am wm
The trace opens in the Perfetto UI when recording finishes.
On Linux, record scheduling activity with tracebox:
curl -LO https://get.perfetto.dev/tracebox && chmod +x tracebox curl -LO https://raw.githubusercontent.com/google/perfetto/main/test/configs/scheduling.cfg sudo ./tracebox -o trace.pftrace --txt -c scheduling.cfg
For Chrome, use the recording page in the UI. To trace your own app, see in-app tracing.
curl -LO https://get.perfetto.dev/trace_processor && chmod +x trace_processor ./trace_processor query trace.pftrace \ "SELECT name, COUNT(*) AS cnt FROM slice GROUP BY name ORDER BY cnt DESC LIMIT 10"
Or from Python:
pip install perfetto
from perfetto.trace_processor import TraceProcessor tp = TraceProcessor(trace='trace.pftrace') print(tp.query('SELECT ts, dur, name FROM slice LIMIT 10').as_pandas_dataframe())
ftrace. (Linux cookbook)chrome://tracing, for debugging the browser, V8 and Blink. (Chrome tracing)Not sure where to start? How do I start using Perfetto? picks a path based on what you want to do.
Perfetto is a collection of components that work together:
traced, traced_probes) collect data from many processes and the kernel into a single trace file.See the documentation for the full picture.
See Community and getting help for more. We follow Google's Open Source Community Guidelines.
Contributions are welcome. The contributor guide covers building, testing and sending pull requests.
The Debian packaging of Perfetto is maintained in the Debian Perfetto Salsa repository.
Apache 2.0. See LICENSE for details.