`co-uring-http`

co-uring-http is a high-performance HTTP server built on C++20 coroutines and io_uring. This project serves as an exploration of the latest features of Linux kernel and is not recommended for production use. For learning purposes, co-uring-http re-implements general-purpose coroutine primitives (such as task<T>, sync_wait<task<T>>, etc.) instead of utilizing existing libraries.

Leverages C++20 coroutines to manage clients and handle HTTP requests, which simplifies the mental overhead of writing asynchronous code.
Leverages io_uring for handling async I/O operations, such as accept(), send(), recv(), and splice(), reducing the number of system calls.
Leverages ring-mapped buffers to minimize buffer allocation costs and reduce data transfer between user and kernel space.
Leverages multishot accept in io_uring to decrease the overhead of issuing accept() requests.
Implements a thread pool to utilize all logical processors for optimal hardware parallelism.
Manages the lifetime of io_uring, file descriptors, and the thread pool using RAII classes.

Design Document

Component

Coroutine

task (task.hpp): The task class represents a coroutine that does not start until it is awaited.
thread_pool (thread_pool.hpp): The thread_pool class provides an abstraction that allows a coroutine to be scheduled on fixed-size pool of threads. The number of threads is limited to the number of logical processors, allowing for hardware parallelism.

File Descriptor

file_descriptor (file_descriptor.hpp): The file_descriptor class owns a file descriptor. The file_descriptor.hpp file provides general-purpose functions that works with the file_descriptor class, such as open(), pipe(), and splice().
server_socket (socket.hpp): The server_socket class extends the file_descriptor class and represents the listening socket that could accept clients. It provides an accept() method, which records if there is an existing multishot accept request in io_uring and submits a new one if none exists.
client_socket (socket.hpp): The client_socket class extends the file_descriptor class and represents the socket that could communicate with the client. It provides a send() method, which submits a send request to io_uring and a recv() method, which submits a recv request to io_uring.

`io_uring`

io_uring (io_uring.hpp): The io_uring class is a thread_local singleton, which owns the submission queue and completion queue of io_uring.
buffer_ring (buffer_ring.hpp): The buffer_ring class is a thread_local singleton, manages a collection of buffers that are supplied to io_uring. When an incoming HTTP request arrives, io_uring selects a buffer from the available pool and populates it with the received data, eliminating the need for allocating a new buffer for each request. Once the http_parser finishes parsing the request, the server returns the buffer to io_uring, enabling its reuse for future requests. The number of buffers and and the size of each buffer are defined in constant.hpp, which can be adjusted based on the estimated workload of the HTTP server.

HTTP Server

http_server (http_server.hpp): The http_server class creates a thread_worker task for each thread in the thread_pool and waits for these tasks to finish. The tasks will finish when an exception occurs.
thread_worker (http_server.hpp): The thread_worker class contains a collection of coroutines that will be scheduled on each thread. When the class is constructed, it spawns thread_worker::accept_client() and thread_worker::event_loop().
- The thread_worker::event_loop() coroutine processes events in the completion queue of io_uring and resumes the coroutine that is waiting for the completion of that event.
- The thread_worker::accept_client() coroutine invokes server_socket::accept() to submit a multishot accept() request to io_uring, which will generate a completion queue event for each incoming client. When a client arrives, it spawns the thread_worker::handle_client() coroutine.
- The thread_worker::handle_client() coroutine invokes client_socket::recv() to wait for an HTTP request. When a request arrives, it parses the request with http_parser (http_parser.hpp), constructs an http_response (http_message.hpp), and sends the response with client_socket::send().

auto thread_worker::handle_client(client_socket client_socket) -> task<> {
  http_parser http_parser;
  buffer_ring &buffer_ring = buffer_ring::get_instance();
  while (true) {
    const auto [recv_buffer_id, recv_buffer_size] = co_await client_socket.recv(BUFFER_SIZE);
    const std::span<char> recv_buffer = buffer_ring.borrow_buffer(recv_buffer_id, recv_buffer_size);

    if (const auto parse_result = http_parser.parse_packet(recv_buffer); parse_result.has_value()) {
      const http_request &http_request = parse_result.value();

      // Processes the `http_request` and constructs an `http_response`
      // with a status that is either `200` or `404`
      // (Please refer to the source code)

      std::string send_buffer = http_response.serialize();
      co_await client_socket.send(send_buffer, send_buffer.size());
    }

    buffer_ring.return_buffer(recv_buffer_id);
  }
}

Workflow

The http_server creates a thread_worker task for each thread in the thread_pool and awaits their completion.
Each thread_worker creates a socket with the SO_REUSEPORT option, allowing the reuse of the same port, and spawns the thread_worker::accept_client() and thread_worker::event_loop() coroutines.
Upon a client arrival, the thread_worker::accept_client() coroutine spawns a thread_worker::handle_client() coroutine to handle HTTP requests for that client.
When either thread_worker::accept_client() or thread_worker::handle_client() awaits an asynchronous I/O operation (such as send() or recv()), it suspends its execution and submits a request to the submission queue of io_uring. Execution control is then transferred back to thread_worker::event_loop().
The thread_worker::event_loop() processes events in the completion queue of io_uring. For each event, it identifies the coroutine that is awaiting that event and resumes its execution.

Performance

The benchmark is performed with the hey benchmark tool, which sends 200 batches of requests, with each batch containing 5,000 concurrent clients requesting a file of 1024 bytes in size. co-uring-http serves 57,012 requests per second and handles 99% of requests within 0.2 seconds.

The benchmark is performed on UTM running on MacBook Air (M1, 2020) with Linux kernel version 6.4.2. The virtual machine has 4 cores and 8 GB memories. The program is compiled with GCC 13 and optimization level O3.

./hey -n 1000000 -c 5000 http://127.0.0.1:8080/1k

Summary:
  Total:        17.5400 secs
  Slowest:      0.3872 secs
  Fastest:      0.0001 secs
  Average:      0.0824 secs
  Requests/sec: 57012.6903

  Total data:   1024000000 bytes
  Size/request: 1024 bytes

Response time histogram:
  0.000 [1]     |
  0.039 [18601] |■
  0.077 [516164]        |■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.116 [344028]        |■■■■■■■■■■■■■■■■■■■■■■■■■■■
  0.155 [89199] |■■■■■■■
  0.194 [22404] |■■
  0.232 [3841]  |
  0.271 [3404]  |
  0.310 [1907]  |
  0.348 [221]   |
  0.387 [230]   |


Latency distribution:
  10% in 0.0519 secs
  25% in 0.0611 secs
  50% in 0.0752 secs
  75% in 0.0973 secs
  90% in 0.1213 secs
  95% in 0.1392 secs
  99% in 0.1922 secs

Details (average, fastest, slowest):
  DNS+dialup:   0.0002 secs, 0.0001 secs, 0.3872 secs
  DNS-lookup:   0.0000 secs, 0.0000 secs, 0.0000 secs
  req write:    0.0001 secs, 0.0000 secs, 0.0966 secs
  resp wait:    0.0413 secs, 0.0000 secs, 0.2908 secs
  resp read:    0.0397 secs, 0.0000 secs, 0.3499 secs

Status code distribution:
  [200] 1000000 responses

Environment

Linux Kernel >= 5.19
GCC >= 13 or Clang >= 14
- Because libc++ lacks support for certain C++20 features such as jthread, co-uring-http should link with GCC's libstdc++.
- There's a bug in GCC <= 12 that introduces unexpected copies in the co_await expression, which will cause a segmentation fault in co-uring-http.
liburing >= 2.3

The .devcontainer/Dockerfile provides a container image based on ubuntu:lunar with the required dependencies installed. Please note that the virtual machine of Docker on macOS is based on Linux kernel 5.15, which doesn't meet the requirement of co-uring-http.

Build

Generate build configuration with CMake:

cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_COMPILER:FILEPATH=/usr/bin/clang -DCMAKE_CXX_COMPILER:FILEPATH=/usr/bin/clang++ -B build -G "Unix Makefiles"

Build and run co_uring_http, which will listen on localhost:8080 and serves static files:

make -C build -j$(nproc)
./build/co_uring_http

Update WSL Kernel

The default Linux kernel version for the Windows Subsystem for Linux (WSL) is 5.15, which does not support certain io_uring features, such as multi-shot accept or ring-mapped buffers. However, WSL provides a .wslconfig file that enables the use of a custom-built kernel image. It's recommended to build the kernel within an existing WSL instance.

Install the build dependencies:

sudo apt install git bc build-essential flex bison libssl-dev libelf-dev dwarves

Download the latest kernel source code:

wget https://github.com/torvalds/linux/archive/refs/tags/v6.4.2.tar.gz -O v6.4.2.tar.gz
tar -xf v6.4.2.tar.gz

Download the build configuration file for WSL:

wget https://raw.githubusercontent.com/microsoft/WSL2-Linux-Kernel/linux-msft-wsl-6.1.y/Microsoft/config-wsl
cp config-wsl linux-6.4.2/arch/x86/configs/wsl_defconfig

Build the kernel:

cd linux-6.4.2
make KCONFIG_CONFIG=arch/x86/configs/wsl_defconfig -j$(nproc)

Clone the kernel image to $env:USERPROFILE (default user path) and set .wslconfig to use the kernel image:

powershell.exe /C 'Copy-Item .\arch\x86\boot\bzImage $env:USERPROFILE'
powershell.exe /C 'Write-Output [wsl2]`nkernel=$env:USERPROFILE\bzImage | % {$_.replace("\","\\")} | Out-File $env:USERPROFILE\.wslconfig -encoding ASCII'

Restart WSL:

wsl --shutdown

xiaoyang-sde/co-uring-http

co-uring-http