aiofastnet 0.8.0

Fast transport/protocol networking for asyncio

aiofastnet

aiofastnet gives your asyncio networking application an instant performance boost, lower latency and higher throughput by just adding two lines:

import aiofastnet

...
# Call this before asyncio.run(...)
aiofastnet.install_policy()

Are you using aiohttp, asyncpg, websockets, uvicorn, or any other library that relies on asyncio networking? They become faster if you enable aiofastnet. The difference is especially noticeable when SSL is used.

How is this possible?

aiofastnet provides drop-in, highly efficient C/Cython replacements for asyncio's:

• loop.create_connection()
• loop.open_connection()
• loop.create_server()
• loop.start_server()
• loop.start_tls()
• loop.sendfile()

asyncio libraries use these loop primitives to establish communication channels. The current implementations in both asyncio and uvloop are far from optimal, especially for SSL/TLS connections.

Calling aiofastnet.install_policy() replaces these primitives with aiofastnet's efficient implementations. From that moment on, any library that uses asyncio networking will use aiofastnet.

aiofastnet is not a different event loop. It works on top of stock asyncio loops or uvloop by using low-level primitives such as add_reader and add_writer. It has no background threads and does not use unscalable tricks such as calling synchronous recv/send syscalls from another thread. Essentially, it provides the same kind of internal implementation you would find in asyncio and uvloop, but with much better optimization.

As a cherry on top, aiofastnet supports Kernel TLS out of the box on Linux.

Benchmark

The benchmark below compares echo round-trips over loopback for TCP and SSL. The exact gains depend on workload, message sizes, CPU, OpenSSL version, and how much of your total runtime is spent in transport/SSL plumbing.

Source: examples/benchmark.py

In these benchmarks, aiofastnet is up to 2.2x faster than standard asyncio and up to 1.6x faster than uvloop for TLS connections.

aiofastnet is fully compatible with free-threaded Python builds and scales as expected when multiple event loops run in parallel across multiple threads.

Source: examples/benchmark_threaded.py

Why Use aiofastnet

• Faster hot path. Transport and SSL internals are reimplemented in Cython/C to reduce overhead on long-lived connections.
• Drop-in API. Keep using the standard asyncio streams or transport/protocol model and familiar loop-level networking operations.
• Works with the event loop you already use. aiofastnet works with stock asyncio loops, uvloop, and winloop.
• Particularly strong for SSL-heavy workloads. aiofastnet uses OpenSSL more directly and avoids extra copies in the data path.
• Kernel TLS support on Linux. Native sendfile for TLS connections through SSL_sendfile. aiohttp will be able to send FileResponse more efficiently over TLS connections.
• Safer transport write() / writelines() behavior. If the socket cannot accept everything immediately, only bytes and memoryview objects backed by bytes are retained without copying. Other objects, including bytearray and non-bytes exporters, are copied before being queued. Unlike standard asyncio transports, this avoids sending corrupted data if application code modifies the underlying buffer after write() / writelines() returns.
• Better SSL backpressure semantics. Buffer sizes and write limits reflect what is actually queued across the stack.
• Works for library authors. WebSocket, HTTP, RPC, proxy, database, broker, and custom protocol libraries can expose the same API while giving users a faster transport layer.

Quickstart

Install from PyPI:

$ pip install aiofastnet

aiofastnet requires Python 3.9 or greater.

For applications, the easiest way to enable aiofastnet is to use an event loop factory.

import asyncio
import aiofastnet

async def main():
    ...

asyncio.run(main(), loop_factory=aiofastnet.loop_factory())

If you use another event loop implementation, pass its loop factory:

import asyncio
import uvloop
import aiofastnet

asyncio.run(main(), loop_factory=aiofastnet.loop_factory(uvloop.new_event_loop))

For older applications that still configure asyncio through event loop policies, install the aiofastnet policy wrapper before creating loops:

import asyncio
import aiofastnet

aiofastnet.install_policy()
asyncio.run(main())

Event loop policies are deprecated in Python 3.14 and are scheduled for removal in Python 3.16, so prefer loop_factory() for new code.

If the event loop already exists, patch it directly:

import asyncio
import aiofastnet

async def main():
    aiofastnet.patch_loop()
    ...

asyncio.run(main())

Library authors should call aiofastnet APIs directly instead of patching a loop owned by their users. The API mirrors stdlib asyncio: pass the running loop and use your existing protocol factory.

import asyncio
import aiofastnet

class EchoClientProtocol(asyncio.Protocol):
    def connection_made(self, transport):
        self.transport = transport
        self.transport.write(b"hello")

    def data_received(self, data):
        print("received:", data)
        self.transport.close()

async def main():
    loop = asyncio.get_running_loop()

    transport, protocol = await aiofastnet.create_connection(
        loop,
        EchoClientProtocol,
        "127.0.0.1",
        9000,
    )

    ...

asyncio.run(main())

For servers, replace loop.create_server(...) with aiofastnet.create_server(loop, ...) in the same way.

aiofastnet also exposes start_tls and sendfile:

transport = await aiofastnet.start_tls(
    loop,
    transport,
    protocol,
    ssl_context,
    server_side=False,
    server_hostname="example.com",
)

await aiofastnet.sendfile(loop, transport, fileobj)

When aiofastnet Is a Good Fit

aiofastnet is most attractive when you already rely on asyncio's transport/protocol APIs directly or indirectly and one or more of these are true:

• You have long-lived TCP or TLS connections.
• You run protocol-heavy services where transport overhead is visible in CPU profiles.
• You maintain a library and want better performance without changing your public API.
• You care about consistent write() / writelines() buffer behavior.
• You want SSL flow control to reflect the whole buffered stack, not only part of it.

When Not To Use aiofastnet

aiofastnet is not the right default for every networking project:

• If your workload is dominated by very short-lived connections, you should expect little or no gain. Currently, aiofastnet focuses on optimizing the data path after connection establishment.

Platform Compatibility

aiofastnet is built and tested on Linux, macOS, and Windows. It works with the standard asyncio event loop, uvloop, and winloop.

On Windows it works with SelectorEventLoop and winloop. With ProactorEventLoop it falls back to asyncio's native connection and server implementation, because the proactor loop does not provide the add_reader() / add_writer() hooks required by aiofastnet's custom transport implementation. For transports created through aiofastnet, a compatibility wrapper preserves the documented write() / writelines() buffer-safety behavior.

Kernel TLS (KTLS) support on Linux

On Linux, aiofastnet can use OpenSSL's KTLS support for TLS connections. KTLS is beneficial if any of the following is true:

• Static files need to be sent over TLS connection and sendfile() can be used. In that case the kernel can read data directly instead of forcing the application to copy file contents through userspace.
• Some high-end NICs support hardware TLS offload. This leads to huge CPU savings.

If you only sent regular data (not static files) and do not have high-end NIC with TLS offload, enabling KTLS may actually cause a slight performance degradation. CPU cost-wise it doesn't matter where encryption/decryption happens in kernel or in userspace, but the kernel tls module has to do extra bookkeeping. Also, aiofastnet can batch data and reduce amount of syscalls when KTLS is not used.

KTLS requires support from all of these layers:

• The tls kernel module loaded.
• OpenSSL built with KTLS support on a machine with suitable kernel headers.
• An ssl.SSLContext with ssl.OP_ENABLE_KTLS enabled.
• A TLS version and cipher suite supported by the kernel TLS implementation.

To load the kernel module:

$ sudo modprobe tls

To enable KTLS on Python SSL contexts (available in Python 3.12+):

import ssl

server_context = ssl.SSLContext(ssl.PROTOCOL_TLS_SERVER)
server_context.options |= ssl.OP_ENABLE_KTLS

client_context = ssl.create_default_context()
client_context.options |= ssl.OP_ENABLE_KTLS

The OpenSSL library used by Python must itself have KTLS support. Python distributions often ship their own libssl and libcrypto, and those libraries may have been built on older systems where KTLS was not available. That can be true even when the host running your program has a newer kernel.

Check which OpenSSL Python is using:

$ python -c "import ssl; print(ssl.OPENSSL_VERSION); print(ssl._ssl.__file__)"

If your Python distribution bundles an older or non-KTLS OpenSSL, one practical option is to locate the bundled libssl and libcrypto files and replace them with symbolic links to a newer system OpenSSL build that supports KTLS. This is often useful with Conda Python, which commonly ships its own OpenSSL libraries inside the environment.

KTLS support by kernel version is outline here.

Check out aiohttp_ktls_fileresponse.py and aiohttp_ws_speedup.py
examples showing how you can speed up aiohttp (or any other asyncio application).

Some other useful links:

• https://dev.to/ozkanpakdil/kernel-tls-nic-offload-and-socket-sharding-whats-new-and-who-uses-it-4e1f
• https://www.kernel.org/doc/html/latest/networking/tls.html

Free-Threaded Python

aiofastnet is compatible with free-threaded Python builds such as python3.14t. The extension modules are built to work without forcing the legacy GIL back on, so separate event loops may run in separate threads.

The repository includes several free-threading examples:

• examples/benchmark_threaded.py runs multiple echo client/server pairs in parallel to compare single-loop and multi-threaded execution.
• examples/echo_server_threaded.py starts one listening echo server per thread on the same port with reuse_port=True so the kernel can distribute incoming connections across worker threads.
• examples/echo_client_threaded.py starts one echo client per thread and drives them in parallel against the shared server port.

Transport objects remain thread-affine. Methods such as write(), writelines(), close(), pause_reading(), and similar transport operations must be called from the same thread that established the connection. Calling transport methods directly from a different thread raises RuntimeError.

This matches the general asyncio model: loops and loop-owned objects are not meant to be used concurrently from arbitrary threads. aiofastnet does not add internal transport locking for cross-thread access.

To use aiofastnet correctly with multithreading:

• Create a separate event loop in each worker thread.
• Create connections and servers inside the loop thread that will own them.
• Keep all direct transport interaction on that same thread.
• If another thread needs to send data or close a transport, schedule that work onto the owning loop with loop.call_soon_threadsafe(...) instead of touching the transport directly.

In other words, free-threaded Python lets multiple loops make progress in parallel, but each individual connection is still owned by exactly one loop and one thread.

Building From Source

1. Clone the repository:

   $ git clone [e-mail]:tarasko/aiofastnet.git
   $ cd aiofastnet
   

2. Create and activate a virtual environment:

   $ python3 -m venv aiofn-dev
   $ source aiofn-dev/bin/activate
   

3. Install test dependencies:

   $ pip install -r requirements-test.txt
   

4. Build extensions in place and run tests:

   $ python setup.py build_ext --dev --inplace
   $ pytest -s -v
   

5. Run the benchmark:

   $ python -m examples.benchmark
   

6. Run tests:

   $ pytest -s -v
   $ pytest -s -v -k test_echo[uvloop-tcp-buffered-6291456] --asyncio-debug --log-cli-level DEBUG
   

7. Build coverage report:

Building for coverage testing requires enabling line tracing in cython, which significantly slows down extension modules. It is disabled by default. You would need to rebuild specifically with coverage support.

python setup.py build_ext --inplace --dev --with-coverage
pytest -s -v --cov=aiofastnet --cov-report=html

Contributing

Contributions are welcome!