Go to file
Fantix King f055323af3
Update gitignore
2022-07-02 18:11:03 -04:00
.github/workflows Add GitHub Action to build 2022-06-26 11:48:48 -04:00
docs Add logo and badge 2022-06-26 11:48:46 -04:00
resolver Update URL in the license text 2022-06-26 11:48:34 -04:00
src/kloop Unify debug setting 2022-07-02 18:08:25 -04:00
tests Add debug script and certs 2022-07-02 18:08:07 -04:00
.gitignore Update gitignore 2022-07-02 18:11:03 -04:00
LICENSE Update URL in the license text 2022-06-26 11:48:34 -04:00
MANIFEST.in Add logo and badge 2022-06-26 11:48:46 -04:00
Makefile Fix setup.py build and clean 2022-06-26 11:48:48 -04:00
README.md Fix README about TLS 1.3 RX offloading 2022-07-02 16:08:49 -04:00
README.zh.md Fix README about TLS 1.3 RX offloading 2022-07-02 16:08:49 -04:00
pyproject.toml PoC uring and ktls 2022-03-20 19:07:05 -04:00
setup.cfg Fix metadata to point to Gitee 2022-07-02 16:10:08 -04:00
setup.py Unify debug setting 2022-07-02 18:08:25 -04:00

README.md

kLoop - asyncio on Linux kernel

中文 build downloads code quality license

kLoop is an implementation of the Python asyncio event loop written in Cython, using io_uring and kTLS features of the Linux kernel, therefore called k(ernel)Loop.

kLoop is open-sourced and released under the MulanPSL - 2.0 license.

⚠️WARNING: THIS PROJECT IS IN PROOF-OF-CONCEPT STAGE!⚠️

Features

  • Minimum syscalls - all I/O calls are done in the kernel thanks to io_uring, and the only remaining syscall to io_uring_enter() is also optimized to be only called when necessary using the SQPOLL feature. That means most of the overhead of syscalls is gone;
  • No GIL in the main-loop - the hot-path is written in Cython without GIL, that means it's compiled into pure C code without Python structures, saving some memory and execution time. On the other hand, GIL is only taken before Python callbacks, so it's slightly more friendly to multithreading, which is still not recommended though.
  • TLS offloading - all symmetric-key encryption and decryption work is offloaded to the NIC if supported, or to the thread pool of io_uring otherwise. This allows us to free up CPU for more I/O, or leverage all the CPU cores even if the application is running single-threaded.
  • Asynchronous DNS resolving - we blended in the Rust trust-dns library with a custom I/O runtime bridging to io_uring in C (including reading the /etc/resolv.conf and /etc/hosts files), providing flexible APIs to manage the concurrency, cache and configuration in Python.

Requirements

  • Python >= 3.10
  • Linux >= 5.11 (enable kTLS with modprobe tls)
  • OpenSSL >= 3.0 (kTLS receive offloading on TLS 1.3 requires the latest development version)

Development and testing is done on Ubuntu 22.04.

Architecture Diagram

architecture.png

Looks like the Lucky Charms factory, says @aaronbrighton ...

Development

Ubuntu 22.04

sudo apt update
sudo apt install gcc libssl-dev python3-dev python3.10-venv
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
python3 -m venv path/to/env
source path/to/env/bin/activate
pip install cython
KLOOP_DEBUG=1 python setup.py develop  # or just run `make`