network performance

[helloldt]

Hello, for the sake of network performance, you should-shield: Info (including Success information) and Warn level logs.

• Keep: Error level logs to ensure that you can see key error information when a program error occurs.
• For other unknown reasons, the current program has network performance problems, which will reduce performance by nearly 60%.

[helloldt]

// Setup logging - check SLOG_LEVEL env var
logLevel := slog.LevelError // Default to error, suppressing info/warn/success
/*
if lvl := os.Getenv("SLOG_LEVEL"); lvl != "" {
switch lvl {
case "debug", "DEBUG":
logLevel = slog.LevelDebug
case "info", "INFO":
logLevel = slog.LevelInfo
case "warn", "WARN":
logLevel = slog.LevelWarn
case "error", "ERROR":
logLevel = slog.LevelError
}
}
*/

[helloldt]

I've been using this project in a LAN environment where multiple clients use the pcap2socks gateway simultaneously. I noticed significant throughput degradation and packet drops during high concurrency or high bandwidth usage.

After analyzing the code, I found that the default buffer sizes and congestion control settings were a bit conservative for modern high-speed networks. I made the following adjustments to optimize performance:

1. Switched to CUBIC Congestion Control :
   Changed tcpCongestionControlAlgorithm from reno to cubic in core/option/option.go . CUBIC handles high-bandwidth and high-latency links much better than Reno.
2. Increased TCP Buffers :
   Increased tcpRXBufMaxSize and tcpTXBufMaxSize from 8MB/6MB to 32MB . This allows the gVisor stack to fully utilize available bandwidth without being throttled by window sizes.
3. Expanded PCAP Capture Buffer :
   Increased the pcap handle buffer size in core/device/pcap.go from 512KB to 8MB . This prevents packet drops at the capture layer during micro-bursts of traffic from multiple clients.
4. Enlarged Outbound Queue :
   Increased defaultOutQueueLen in core/device/iobased/endpoint.go from 1024 to 4096 to reduce packet drops in the internal processing queue under heavy load.
   These changes have significantly improved stability and speed in my multi-client setup. I hope this feedback is helpful!

Best regards.