kepler 0.3.0

Fast, easy, pretty and readable program instrumentation.

Kepler

Wish you could see stuff like this about your program?

Kepler meticulously tracks your program, and creates simply and easily readable reports to help you understand what they're doing.

Kepler is not a replacement for a good profiling tool, nor is it necessarily a great production implementation tool. Kepler is designed to be that go-to tool in your toolbelt for quick and dirty measurements of your programs.

Installing Kepler

pip install kepler

Kepler in action

The first thing you should do with Kepler is annotate a function or two you want to time with @kepler.time, and then add a kepler.report() call to your amin function.

Here's the script that produced the screenshot above:

import kepler, random, time

@kepler.time("do some stuff")
def do_some_stuff():
    split = kepler.stopwatch("watch")
    for i in kepler.time("loop", range(20)):
        with kepler.time("sleep"):
            time.sleep(random.random() / 100)
        if i % 2 == 1:
            with kepler.time("overhead"):
                split("odd")
        else:
            with kepler.time("overhead"):
                split("even")

@kepler.time
def main():
    with kepler.time("sloooow"):
        time.sleep(0.3)
    do_some_stuff()

main()
kepler.report()

Adding custom timers

Custom timers may be implemented with the measurement decorator. This works similarly to Python's contextlib.contextmanager decorator:

• The decorated function should be a generator
• It should do any setup it needs to do, and then yield exactly once
• yield call corresponds exactly to the code to be measured
• It should return an Event to be added to kepler's log

import time
import kepler
import torch

@kepler.measurement
def time_gpu():
    start_time = time.time_ns()
    start, end = torch.cuda.Event(True), torch.cuda.Event(True)
    start.record()
    yield
    end.record()
    torch.cuda.synchronize()
    return kepler.TimingEvent(start_time, start.elapsed_time(end) * 1e6)


with time_gpu("matmul"):
    _ = torch.rand([2, 2]) @ torch.rand([2, 2])

kepler.report()

Roadmap

✅ Changelog

• Custom timing measurements
• Import and export reports as json
• Report directly from json -- try python -m kepler.report < tests/data/simple_log.json
• Testing for units, json import and export

🔜 Up next

• Track system metrics

🌈 Before 1.0

• Export traces to pandas
• Flamegraphs
• Integrate with open-telemetry
• Track and report other metrics besidings timings
• Thorough unit testing
• Docs
• Examples
• Logo