Scheduler benchmark framework, written in Rust.
Currently implements two tests simulating two important workloads for Meta production: web and xdb.
Tests are in src/tests. Look at the web test for the example on how to write tests.
Build/run in the --release mode, otherwise it will take too long for tests to run.
Results might be flaky if the test is runned on a laptop due to the cpu frequency scaling and power throttling.
Tests are running until a stable result will be obtained or the number of attempts will exceed 10.
E.g.:
n 1 1519353
n 2 1580586 (± 16.4%)
--------------- discard ----------------
n 1 1505764
n 2 1514584 (± 2.5%)
n 3 1544058 (± 10.1%)
--------------- discard ----------------
n 1 1617176
n 2 1606153 (± 2.9%)
n 3 1645072 (± 12.5%)
--------------- discard ----------------
n 1 1732681
n 2 1738164 (± 1.3%)
n 3 1746761 (± 2.7%)
n 4 1749062 (± 2.4%)
n 5 1776681 (± 10.6%)
--------------- discard ----------------
n 1 1812608
n 2 1826930 (± 3.3%)
--------------- discard ----------------
n 1 1826546
n 2 1850692 (± 5.5%)
--------------- discard ----------------
n 1 1792598
n 2 1834142 (± 9.6%)
--------------- discard ----------------
n 1 1892144
n 2 1869416 (± 5.2%)
--------------- discard ----------------
n 1 1980341
n 2 1980884 (± 0.1%)
========================================
Result: 1.980884s ± 2.305ms (± 0.1%)
Usage:
-
Run web test:
cargo run --release run web -
Run xdb test:
cargo run --release run xdb -
Get data to calibrate compute() cost (takes a while):
cargo run --release calibrate