Profiling Tremor¶

This is a short canned synopsis of profiling tremor.

Valgrind¶

We use valgrind, specifically callgrind, to drive call graph optimizations in tremor.

Install valgrind via homebrew

brew install valgrind

The Google Performance Toolkit, gprof2dot, qcachegrind are also required / useful.

brew install gprof2dot
brew install gperftools
brew install qcachegrind

This is good enough for initial high-level exploration.

For example, run a tremor pipeline against recorded data in data.json

valgrind --tool=callgrind target/debug/tremor run tests/configs/ut.combine3-op.yaml data.json

Analysing results via google perf toolkit and graphviz for static call flow diagrams

gprof2dot -f callgrind callgrind.out.93972 > pipe.dot
dot -Tpng pipe.dot -o pipe.png && open pipe.png

Interactive analysis via QCachegrind / KCachegrind

qcachegrind callgrind.out.93972

The profiling ( sampling ) frequency is tunable and SHOULD be tuned for each run, eg:

RUST_BACKTRACE=1 PROFILEFREQUENCY=1000 valgrind --tool=callgrind \
    target/release/tremor run examples/config-spike5.tremor data.json

Note

When using a release build make sure debug symbols are configured in Cargo.toml and enable link time optimisations ( LTO ).

Install rust flamegraph support:

cargo install flamegraph

Perform a benchmark run with flamegraph support ( on Mac OS X ):

flamegraph target/release/tremor server run -f bench/real-workflow-througput-json.yaml bench/link.yaml

This generates a flamegraph.svg file which can be opened from a browser.

open flamegraph.svg