1. Home
2. Docs
3. Performance

Performance

Benchmark Results

Benchmarks compare Yamux against raw TCP, the reference Go implementation, and Nerdbank.Streams.

Methodology

• 32KB chunk size for all tests
• Parameters: 1/50/500 MB total data, 1/5/20 concurrent streams
• TCP benchmarks use loopback (127.0.0.1)
• In-memory benchmarks use System.IO.Pipelines
• Go comparison uses the HashiCorp Yamux reference implementation (v0.1.1)
• Nerdbank.Streams uses protocol version 1

Expected Results

Benchmarks should show:

1. SocketBaselineAsync — Raw TCP throughput (theoretical maximum)
2. YamuxStreamAsync — .NET Yamux over TCP with multi-stream parallelism
3. CsharpToGoAsync — C# Yamux client vs Go Yamux server (interop benchmark)
4. NerdbankStreamsAsync — Nerdbank.Streams over in-memory streams
5. YamuxInMemoryAsync — Yamux over in-memory pipes (theoretical max)

Running Benchmarks

dotnet run -c Release --project benchmarks/Yamux.Benchmark

Allocation Characteristics

The library is designed to minimize GC pressure:

• ReusableValueTaskSourcePool pools TaskCompletionSource instances used for write completion signaling
• System.IO.Pipelines provides zero-copy buffer management for reading
• Frame struct is stack-allocated; payload ownership is tracked via IDisposable
• StreamIdGenerator uses Interlocked.Add for lock-free ID generation
• RemoteDataWindow uses pooled waiter objects for async flow control

Throughput Tuning

• MaxDataFrameSize — Larger frames reduce per-frame overhead but increase latency. Default 16KB.
• ReceiveWindowSize — Larger windows improve throughput for high-BDP connections. Default 256KB.
• AutoTuneReceiveWindowSize — Automatically scales window up to ReceiveWindowUpperBound based on RTT.
• ReceiveWindowUpperBound — Caps auto-tuning. Default 16MB.