zerg

Low-level TCP server framework for C# built on Linux io_uring. Direct control over sockets, buffers, and scheduling with no hidden abstractions.

Requirements: Linux kernel 6.1+, .NET 8/9/10.

Install

dotnet add package zerg

Quick Start

using zerg.Engine;
using zerg.Engine.Configs;

var engine = new Engine(new EngineOptions { Port = 8080, ReactorCount = 1 });
engine.Listen();

while (engine.ServerRunning)
{
    var connection = await engine.AcceptAsync(CancellationToken.None);
    if (connection is null) continue;
    _ = HandleAsync(connection);
}

static async Task HandleAsync(Connection connection)
{
    while (true)
    {
        var result = await connection.ReadAsync();
        if (result.IsClosed) break;

        var rings = connection.GetAllSnapshotRingsAsUnmanagedMemory(result);
        // process rings.ToReadOnlySequence() ...
        rings.ReturnRingBuffers(connection.Reactor);

        connection.Write("HTTP/1.1 200 OK\r\nContent-Length: 2\r\n\r\nOK"u8);
        await connection.FlushAsync();
        connection.ResetRead();
    }
}

Read API

// High-level: get all buffers as a ReadOnlySequence
var result = await connection.ReadAsync();
var rings = connection.GetAllSnapshotRingsAsUnmanagedMemory(result);
ReadOnlySequence<byte> seq = rings.ToReadOnlySequence();
rings.ReturnRingBuffers(connection.Reactor);
connection.ResetRead();

// Low-level: consume one buffer at a time
while (connection.TryGetRing(result.TailSnapshot, out RingItem ring))
{
    ReadOnlySpan<byte> data = ring.AsSpan();
    connection.ReturnRing(ring.BufferId);
}
connection.ResetRead();

Adapters:

// Zero-copy PipeReader (buffers held until AdvanceTo)
var reader = new ConnectionPipeReader(connection);
var result = await reader.ReadAsync();
reader.AdvanceTo(consumed, examined);

// BCL Stream (one copy per read)
var stream = new ConnectionStream(connection);
int n = await stream.ReadAsync(buffer);

Write API

connection.Write("data"u8);
await connection.FlushAsync();

// Or via IBufferWriter<byte>
Span<byte> span = connection.GetSpan(256);
connection.Advance(bytesWritten);
await connection.FlushAsync();

Configuration

var engine = new Engine(new EngineOptions
{
    Port = 8080,
    ReactorCount = 4,
    AcceptorConfig = new AcceptorConfig(IPVersion: IPVersion.IPv6DualStack),
    ReactorConfigs = Enumerable.Range(0, 4).Select(_ => new ReactorConfig(
        RecvBufferSize: 32 * 1024,
        BufferRingEntries: 16 * 1024,
        IncrementalBufferConsumption: false  // set true for kernel 6.12+
    )).ToArray()
});

Key ReactorConfig options:

Option	Default	Description
RingEntries	8192	io_uring SQ/CQ depth
RecvBufferSize	32KB	Per-buffer size
BufferRingEntries	16384	Number of pre-allocated recv buffers
BatchCqes	4096	Max CQEs per loop iteration
CqTimeout	1ms	Wait timeout (nanoseconds)
IncrementalBufferConsumption	false	Per-connection buffer rings (kernel 6.12+)

Architecture

One acceptor thread distributes connections round-robin to N reactor threads. Each reactor owns its own io_uring instance, buffer ring, and connection map. No locks on hot paths — all cross-thread coordination uses lock-free MPSC queues.

Key features: multishot accept/recv, provided buffer rings, DEFER_TASKRUN, SINGLE_ISSUER, optional SQPOLL, zero-allocation async via IValueTaskSource, connection pooling.

Examples

dotnet run --project Examples -- raw          # zero-copy ring API
dotnet run --project Examples -- pipereader   # PipeReader adapter
dotnet run --project Examples -- stream       # Stream adapter
dotnet run --project Examples -- sqpoll       # SQPOLL mode

Project Structure

core/           Shared library (utils, ConnectionBase, adapters)
zerg/           Main library (Engine, Reactor, native io_uring shim)
terraform/      Alternative pure-C# io_uring implementation (no native deps)
Examples/       Usage examples
Tests/          End-to-end tests

License

MIT