hwcounter

Highly accurate counter for measuring elapsed time in Python


Keywords
benchmark, x86, rdtsc, timing, benchmarking, python, python3
License
Apache-2.0
Install
pip install hwcounter==0.1.0

Documentation

hwcounter

Highly accurate counter for measuring elapsed time in Python.

Installation

$ pip install hwcounter

Overview

This Python extension module uses the hardware timestamp counter to provide very high precision and accurate measurements of execution time.

The module exposes three main objects. Two are plain functions, count and count_end, which return the current value of the timestamp counter. They can be used together to manually track the start and end of a timing measurement. The third is the Timer type, which is to be used as a context manager, wrapping a block of code to be measured.

Using Timer to wrap a block of code is to be preferred over manually using count and count_end. This is for convenience, as well as because it factors in the overhead of calling the underlying measurement instructions.

count()

Returns the current value of the timestamp counter, in cycles.

count_end()

Returns the current value of the timestamp counter, in cycles. It is suitable for use at the end of a timing measurement.

Timer

Class that implements the context manager protocol.

Timer.cycles

The cycles attribute is populated with the elapsed time, in cycles, when the Timer context manager exits.

Example usage

from hwcounter import Timer, count, count_end

from time import sleep
from math import sqrt


# 1. Manually count cycles elapsed between two points

start = count()
sqrt(144) / 12
elapsed = count_end() - start
print(f'elapsed cycles: {elapsed}')
# elapsed cycles: 36486


# 2. Use Timer object as context manager to wrap a block of code and measure its timing

with Timer() as t:
	sleep(1)
print(f'elapsed cycles: {t.cycles}')
# elapsed cycles: 2912338344

These examples were performed on an Intel Core i5-6267U CPU @ 2.90GHz. Notice that the sleep for 1 second in the example above yields a 2.9 billion cycle count.

Notes

The overhead of calling the underlying measurement instructions is taken into account when using the Timer context manager. In other words, the number of cycles it takes to call the machine instructions are subtracted from the elapsed cycle count automatically.

This library returns measurements in processor clock cycles. For benchmarking programs and making apples-to-apples comparisons of changes in code execution time, this method is sufficient and reliable. If elapsed time in seconds is desired, a conversion from clock cycles is required: divide the cycle count by the processor's clock speed (in Hz). This conversion is outside the scope of this module.

Portability

This extension uses the RDTSC and RDTSCP instructions on the x86 architecture, so it won't work on other platforms. It is Python 3 only.