Split View: 비동기 I/O 모델 완전 가이드 2025: epoll, io_uring, Reactor/Proactor, async/await 내부

Complete Guide to Async I/O Models 2025: epoll, io_uring, Reactor/Proactor, and async/await Internals

TL;DR

• Async I/O is the foundation of modern servers: Nginx, Node.js, Redis — every high-performance server.
• Evolution: select → poll → epoll/kqueue → io_uring.
• Reactor vs Proactor: event readiness notification vs operation-completion notification.
• io_uring: the Linux 5.1+ revolution. More efficient than epoll, and truly async.
• async/await: the compiler rewrites your code as a state machine.

1. Evolution of I/O Models

1.1 Five I/O Models

POSIX classification:

1. Blocking I/O: wait until the operation completes.
2. Non-blocking I/O: return immediately, poll for completion.
3. I/O Multiplexing: select/poll/epoll.
4. Signal-driven I/O: notified via signals.
5. Asynchronous I/O: true async (POSIX AIO, io_uring).

1.2 Blocking I/O — The Simplest

data = sock.recv(1024)  # blocks until data arrives
process(data)

Problems:

• One thread = one connection.
• 10K connections = 10K threads (memory blow-up).
• The C10K problem.

1.3 Multi-Threaded as the Fix?

def handle_client(sock):
    data = sock.recv(1024)
    process(data)

while True:
    client = server.accept()
    threading.Thread(target=handle_client, args=(client,)).start()

Limits:

• Per-thread memory (1–8 MB).
• Context-switching cost.
• 10K threads = death.

1.4 Non-blocking I/O

sock.setblocking(False)
try:
    data = sock.recv(1024)
except BlockingIOError:
    pass  # no data — do something else

Problem: infinite-loop polling pins CPU at 100%.

1.5 Enter I/O Multiplexing

Idea: one syscall watching many fds at once.

ready, _, _ = select([sock1, sock2, sock3], [], [])
for sock in ready:
    data = sock.recv(1024)

A single thread can now handle thousands of connections.

2. select and poll

2.1 select (1983)

fd_set readfds;
FD_ZERO(&readfds);
FD_SET(sock, &readfds);

int n = select(max_fd + 1, &readfds, NULL, NULL, &timeout);

if (FD_ISSET(sock, &readfds)) {
    // readable
}

Problems:

• FD_SETSIZE limit (typically 1024).
• O(n) scan: every call inspects every fd.
• fd_set copy: user space ↔ kernel on each call.

2.2 poll (1986)

struct pollfd fds[1024];
fds[0].fd = sock;
fds[0].events = POLLIN;

int n = poll(fds, 1024, timeout);

if (fds[0].revents & POLLIN) {
    // readable
}

Improvements:

• No hard fd-count limit.
• Dynamic array.

Still problematic:

• O(n) scan.
• The fd list is passed on every call.

2.3 Limits of select/poll

10,000 connections with only 100 active:

• select/poll scan all 10,000 every time.
• Wasted CPU.

A new mechanism was needed.

3. epoll — The Linux Revolution

3.1 Arrival (2002, Linux 2.5.44)

int epfd = epoll_create1(0);

struct epoll_event ev;
ev.events = EPOLLIN;
ev.data.fd = sock;
epoll_ctl(epfd, EPOLL_CTL_ADD, sock, &ev);

struct epoll_event events[64];
int n = epoll_wait(epfd, events, 64, -1);

for (int i = 0; i < n; i++) {
    int fd = events[i].data.fd;
    // handle
}

3.2 Why epoll Wins

1. O(1) event detection:

• Kernel returns only ready fds.
• 10K connections + 100 active returns 100.

2. Register once:

• epoll_ctl registers the fd.
• No need to pass them every call.

3. Trigger modes:

• Level-Triggered (LT): keep notifying while data is present (default).
• Edge-Triggered (ET): notify only on state change (Nginx).

3.3 ET vs LT

LT (Level-Triggered):

// 16KB available, you read 8KB → next epoll_wait still notifies

Easier to use, slightly slower.

ET (Edge-Triggered):

// Only notified when new data arrives → must drain fully
while (true) {
    n = recv(fd, buf, sizeof(buf), 0);
    if (n < 0 && errno == EAGAIN) break;  // fully drained
    process(buf, n);
}

Faster, but easier to lose events if you mishandle.

3.4 Systems Built on epoll

• Nginx — core.
• Node.js (via libuv).
• Redis.
• HAProxy.
• Memcached.
• Nearly every high-performance Linux server.

3.5 Equivalents on Other OSes

4. kqueue (BSD/macOS)

4.1 More Powerful Than epoll

int kq = kqueue();

struct kevent change;
EV_SET(&change, sock, EVFILT_READ, EV_ADD | EV_ENABLE, 0, 0, NULL);
kevent(kq, &change, 1, NULL, 0, NULL);

struct kevent events[64];
int n = kevent(kq, NULL, 0, events, 64, NULL);

Richer than epoll:

• File changes (EVFILT_VNODE).
• Signals (EVFILT_SIGNAL).
• Timers (EVFILT_TIMER).
• Process exit (EVFILT_PROC).
• Disk I/O, and more.

Downside: smaller user base than epoll.

5. io_uring — The New Revolution

5.1 Limits of epoll

Even epoll has limits:

• Synchronous I/O: you still call read/write after epoll_wait.
• Syscall cost: user space ↔ kernel on every call.
• Not truly async: read/write themselves can still block.

5.2 io_uring (2019, Linux 5.1)

Built by Jens Axboe (Linux kernel developer). True async I/O.

Core ideas:

• Two ring buffers (Submission, Completion).
• Shared memory between user space and kernel.
• Near-zero syscalls.

[User Space]                    [Kernel Space]
[Submission Queue (SQ)] ────→ [Worker]
                                ↓
[Completion Queue (CQ)] ←──── [I/O complete]

5.3 Example Usage

struct io_uring ring;
io_uring_queue_init(256, &ring, 0);

// Submit work
struct io_uring_sqe *sqe = io_uring_get_sqe(&ring);
io_uring_prep_read(sqe, fd, buf, sizeof(buf), 0);
io_uring_submit(&ring);

// Wait for completion
struct io_uring_cqe *cqe;
io_uring_wait_cqe(&ring, &cqe);
// cqe->res = bytes read
io_uring_cqe_seen(&ring, cqe);

5.4 Advantages of io_uring

1. Truly async:

• Disk I/O no longer blocks.
• Unified model for network and disk.

2. Fewer syscalls:

• Batch submission.
• Some modes allow zero syscalls.

3. Many operations:

• read, write, send, recv.
• accept, connect.
• fsync, fallocate.
• splice, tee.
• statx, openat, close.

4. Faster:

• 30–50% improvement over epoll on certain workloads.

5.5 io_uring in the Wild

• Nginx 1.21+: optional support.
• PostgreSQL 17: partial adoption.
• Tokio (Rust): io-uring backend.
• ScyllaDB: core technology.
• Cloud Hypervisor.

5.6 Future of io_uring

• eBPF integration.
• Direct NVMe access.
• GPU/FPGA I/O.
• All syscalls async.

io_uring is the future of Linux I/O.

6. Reactor vs Proactor

6.1 Reactor Pattern

[Event Loop]
   ↓ epoll_wait
[Event: fd=5 readable]
   ↓
[Handler]
   ↓
[read(5, ...)]  ← the user calls read directly

Characteristics:

• Event-readiness notification.
• The user calls read/write.
• Built on epoll, kqueue.

Examples: Nginx, Node.js, Redis.

6.2 Proactor Pattern

[User] → "read 1024 bytes from this fd"
   ↓
[Kernel] → async work
   ↓
[Completion: data ready]
   ↓
[Handler] → user receives data

Characteristics:

• Delegate the operation itself to the kernel.
• Notified on completion.
• Built on IOCP, io_uring.

Examples: Boost.ASIO, Windows IOCP, io_uring.

6.3 Comparison

6.4 ASIO's Unified Model

Boost.ASIO (C++) exposes Reactor and Proactor through the same interface:

async_read(socket, buffer, [](error_code ec, size_t bytes) {
    // called on completion
});

Internally: epoll on Linux, IOCP on Windows, io_uring on recent builds.

7. async/await Internals

7.1 What async/await Actually Does

async def fetch_data():
    data = await http_request("https://api.example.com")
    return process(data)

There is no magic. The compiler rewrites this as a state machine.

7.2 The Rewrite (Pseudo-code)

class FetchDataStateMachine:
    state = 0

    def resume(self):
        if self.state == 0:
            self.future = http_request("https://api.example.com")
            self.future.set_callback(self.resume)
            self.state = 1
            return  # suspend

        if self.state == 1:
            data = self.future.result()
            return process(data)

Core idea:

• Suspend the function at each await.
• Resume when the result is ready.
• Yield the stack back to the event loop.

7.3 The Event Loop

loop = asyncio.get_event_loop()

while True:
    # 1. Run tasks on the ready queue
    while ready_queue:
        task = ready_queue.pop()
        task.run()

    # 2. Wait for I/O events via epoll_wait
    events = epoll.wait(timeout)

    # 3. Move waiters back onto the ready queue
    for event in events:
        task = pending[event.fd]
        ready_queue.append(task)

7.4 Python asyncio

import asyncio

async def main():
    # Concurrent execution
    results = await asyncio.gather(
        fetch_user(1),
        fetch_user(2),
        fetch_user(3),
    )
    return results

asyncio.run(main())

Internals:

• The selectors module (abstracts epoll/kqueue/select).
• Event loop.
• Tasks and Futures.
• Coroutines.

7.5 JavaScript (Node.js)

async function main() {
    const data = await fetch("https://api.example.com")
    const json = await data.json()
    console.log(json)
}

Internals:

• libuv (C library).
• Event loop.
• epoll/kqueue/IOCP.

7.6 Rust Tokio

#[tokio::main]
async fn main() {
    let data = fetch("https://api.example.com").await;
    println!("{:?}", data);
}

Internals:

• mio (abstracts epoll/kqueue).
• Optional io-uring.
• M:N scheduling across multiple OS threads.
• Work-stealing.

8. Single-Threaded vs Multi-Threaded

8.1 Single-Threaded (Node.js, Redis, Nginx)

[1 thread]
   ↓
[Event Loop] (epoll)
   ↓
[Callback 1] [Callback 2] [Callback 3]

Pros:

• No locks (no race conditions).
• Simple.
• Memory-efficient.

Cons:

• CPU work blocks the loop.
• Only one core is used.

8.2 Multi-Threaded + Event Loops

[N threads (one per CPU core)]
   ↓
[Event Loop per thread]
   ↓
[Work-stealing]

Examples: Tokio (Rust), Akka (JVM), Kotlin Coroutines.

Pros:

• Multi-core utilization.
• Hundreds of thousands of concurrent tasks in a single process.

Cons:

• Synchronization required.
• Harder to debug.

8.3 SO_REUSEPORT

Linux 3.9+ feature:

setsockopt(sock, SOL_SOCKET, SO_REUSEPORT, &one, sizeof(one));

Multiple processes can listen on the same port:

• The kernel load-balances automatically.
• Each process gets its own epoll instance.
• Used by Nginx and HAProxy.

[Port 80]
   ├─ [Worker 1] (epoll)
   ├─ [Worker 2] (epoll)
   └─ [Worker 3] (epoll)

9. Performance Comparison

9.1 Handling 10K Connections

9.2 Real Benchmarks

HTTP server (10K connections):

Apache (prefork):    5,000 req/s
Apache (worker):     20,000 req/s
Nginx (epoll):       100,000 req/s
Nginx (io_uring):    150,000 req/s

ScyllaDB (io_uring):

• 10x+ throughput over Cassandra.
• 1M+ ops/s on a single node.

9.3 NewSQL and Storage

ScyllaDB's secret sauce:

• io_uring + DPDK.
• One shard per core.
• Lock-free.
• Userspace TCP stack.

10. In Practice — Building a High-Performance Server

10.1 Python (asyncio)

import asyncio

async def handle_client(reader, writer):
    while True:
        data = await reader.read(1024)
        if not data:
            break
        writer.write(data)
        await writer.drain()
    writer.close()

async def main():
    server = await asyncio.start_server(handle_client, '0.0.0.0', 8080)
    async with server:
        await server.serve_forever()

asyncio.run(main())

Internally uses selectors → epoll.

10.2 Rust (Tokio)

use tokio::net::TcpListener;
use tokio::io::{AsyncReadExt, AsyncWriteExt};

#[tokio::main]
async fn main() {
    let listener = TcpListener::bind("0.0.0.0:8080").await.unwrap();

    loop {
        let (mut socket, _) = listener.accept().await.unwrap();

        tokio::spawn(async move {
            let mut buf = [0; 1024];
            loop {
                let n = socket.read(&mut buf).await.unwrap();
                if n == 0 { return }
                socket.write_all(&buf[..n]).await.unwrap();
            }
        });
    }
}

Tokio handles:

• mio (epoll).
• M:N scheduling.
• Work-stealing.

10.3 Go

listener, _ := net.Listen("tcp", ":8080")

for {
    conn, _ := listener.Accept()
    go handle(conn)
}

func handle(conn net.Conn) {
    buf := make([]byte, 1024)
    for {
        n, err := conn.Read(buf)
        if err != nil { return }
        conn.Write(buf[:n])
    }
}

The Go runtime handles:

• netpoll (epoll/kqueue).
• Goroutine scheduling.
• Channel integration.

10.4 Comparison

Quiz

References

• The C10K Problem — Dan Kegel
• epoll man page
• Introduction to io_uring — Jens Axboe
• tokio — Rust async runtime
• Boost.ASIO
• libuv — the foundation of Node.js
• Nginx Event Loop
• What is io_uring?
• ScyllaDB — powered by io_uring
• The Reactor Pattern
• Python asyncio Internals