A high-performance Rust implementation of the SWIM gossip protocol using mio and Linux epoll.
┌─────────────────────────────────────────────────────────────────────────────┐
│ Node 1 (seed) Node 2 Node 3 │
│ 127.0.0.1:9000 127.0.0.1:9001 127.0.0.1:9002 │
├─────────────────────────────────────────────────────────────────────────────┤
│ │
│ [12:35:05] Node started [12:35:05] Joining... │
│ [12:35:05] ← PING from :9001 [12:35:05] PING → :9000 │
│ [12:35:05] ACK → :9001 [12:35:05] ← ACK (RTT: 144µs) ✓ │
│ │
│ === TICK === === TICK === === TICK === │
│ Members: 2 active Members: 1 active Members: 1 │
│ RTT: 64µs mean, 7µs jitter RTT: 57µs mean │
│ │
│ [12:35:15] Kill Node 2 (Ctrl+C) [TERMINATED] │
│ │
│ [12:35:16] PING → :9001 ... │
│ [12:35:16] timeout! trying indirect probe │
│ [12:35:17] ⚠ Member :9001 is now SUSPECT │
│ [12:35:20] ✗ Member :9001 is now DEAD │
│ │
└─────────────────────────────────────────────────────────────────────────────┘
Try it yourself:
just cluster # Start 4-node cluster, then type: kill 2Measured on localhost with strace and protocol-level metrics:
| Metric | Value |
|---|---|
| RTT (ping → ack) | 46-144 µs |
| Mean latency | ~60-90 µs |
| P99 latency | ~144 µs |
| Jitter | 7-33 µs |
| Idle CPU | 0% (epoll blocks efficiently) |
epoll_wait(...) = 1 <0.000010s> ← event ready
sendto(9 bytes) <0.000052s> ← send ping
recvfrom(9 bytes) <0.000014s> ← receive ack
epoll_wait(...) = 0 <1.001033s> ← sleep 1s (zero CPU!)
| poll() / select() | epoll() |
|---|---|
| O(n) - scan ALL fds | O(1) - only ready fds |
| Copy fd set every call | Register once, reuse |
| 10k connections = 10k checks | 10k connections = check only active |
# Install
git clone https://github.com/Paulius0112/swim-rs
cd swim-rs
cargo build --release
# Run 4-node cluster
just cluster
# Or manually in separate terminals:
just node1 # Seed node on :9000
just node2 # Joins via :9000
just node3
just node4 ┌──────────┐ PING ┌──────────┐
│ Node A │ ───────────────────► │ Node B │
│ │ ◄─────────────────── │ │
└──────────┘ ACK └──────────┘
│
│ timeout?
▼
┌──────────┐ PING-REQ ┌──────────┐
│ Node A │ ───────────────────► │ Node C │
│ │ "ping B for me" │ │
└──────────┘ └──────────┘
│
│ PING
▼
┌──────────┐
│ Node B │
│ (dead?) │
└──────────┘
State Machine:
Active ──(probe timeout)──► Suspect ──(suspect timeout)──► Dead
▲ │
└────────(ack received)──────┘
| Constant | Value | Description |
|---|---|---|
TICK_INTERVAL |
1s | How often to probe members |
PROBE_TIMEOUT |
500ms | Time to wait for Ack |
SUSPECT_TIMEOUT |
3s | Time before Suspect → Dead |
INDIRECT_PROBE_COUNT |
3 | Nodes to ask for indirect probe |
just bench # Run 30s benchmark
just trace-syscalls 127.0.0.1:9000 # strace epoll/sendto/recvfrom
just perf-stat 127.0.0.1:9000 # CPU performance counters
just flamegraph 127.0.0.1:9000 # Generate flamegraph
just visualize results/*.log # Plot latency chartssrc/
├── main.rs # CLI entry point
├── lib.rs # Library exports
└── protocol/
├── node.rs # Core Node implementation + event loop
├── messages.rs # Ping, Ack, PingReq
└── metrics.rs # RTT tracking, jitter calculation
bench/
├── benchmark.sh # Run cluster and collect stats
├── trace_syscalls.sh # strace wrapper
├── analyze_trace.py # Parse strace output
└── visualize_latency.py # Plot RTT distribution
- SWIM Paper - Original protocol by Das, Gupta, Motivala
- mio - Metal I/O library for Rust
- epoll(7) - Linux I/O event notification
MIT