Transport Layer Performance Improvements

[iduartgomez]

Transport Layer Performance Improvements

Background

PR #2224 added performance benchmarks and analysis for the transport layer. This issue tracks implementing the identified optimizations.

Benchmark Results Summary

Component	Time	% of Packet Path
UDP syscall (send)	12.7µs	70%
Serialization	3.82µs	21%
AES-GCM encryption	1.14µs	6%
Nonce generation	34.7ns	<1%

Key finding: Channel buffer size of 1 vs 100 showed 27x throughput difference.

Proposed Improvements

High Priority (Low Effort)

• Increase channel buffer sizes - Change channel(1) to channel(64) or higher in packet routing paths. This is the highest-impact, lowest-risk change.

• Switch to counter-based nonces - Replace rand::random() nonce generation with atomic counter + connection ID. 5.5x faster (34.7ns → 6.3ns).

Medium Priority (Medium Effort)

• Implement syscall batching - Use sendmmsg/recvmmsg to batch multiple packets per syscall. Potential 5-10x improvement in packet throughput. Requires platform-specific code paths.

• Pre-allocate packet buffers - Pool packet buffers instead of allocating per-packet to reduce allocation overhead.

Lower Priority (Higher Effort)

• Evaluate zero-copy paths - Investigate bytes::Bytes or similar for avoiding copies between encryption and send.

• Alternative serialization - Benchmark rkyv or postcard as bincode alternatives (serialization is 21% of packet creation).

Running Benchmarks

cargo bench --bench transport_perf --no-default-features --features "redb,websocket"

# Or specific levels
cargo bench --bench transport_perf -- level0  # Pure computation
cargo bench --bench transport_perf -- level1  # Mock I/O (channel tests)
cargo bench --bench transport_perf -- level2  # Real sockets

[github-actions]

🤖 Auto-labeled

Applied labels:

• T-enhancement (95% confidence)
• E-medium (80% confidence)
• A-networking (95% confidence)

Reasoning: The issue describes planned performance optimizations to the transport layer (not a bug or new feature), so T-enhancement fits. Work spans small low-effort changes (channel buffer, nonce) and medium/higher-effort items (syscall batching, zero-copy, alternate serialization), so overall effort is estimated as E-medium. Primary areas affected are networking (packet routing, syscalls, buffering) and some cryptography (nonce generation, AES-GCM), hence A-networking and A-crypto.

Previous labels: none

If these labels are incorrect, please update them. This helps improve the auto-labeling system.