This is a benchmark that compares a few different basic TCP proxy implementations in Go and Rust.
prepare-and-run.sh builds and executes the benchmark and it requires Go and Rust toolchains.
The benchmark measures the latency of the proxy implementations by sending HTTP requests over a TCP connection repeatedly.
There are two test cases:
- Make a HTTP/1 GET request to /test1, which returns a 64K-long response body.
- Make a HTTP/1 GET request to /test2, which returns a simple "Hello, world!" response.
Below is the output of a sample run on Ubuntu 20.04 running on AMD Ryzen 9 3950X.
- Go version: 1.16.4
- Rust version: 1.52.1
- Tokio version: 1.6.0
The Go implementation uses io.Copy which has an 32K internal buffer. The std and Tokio-based Rust implementations offer a command-line argument to set the buffer size.
| Test | Case | Time | Throughput |
|---|---|---|---|
| Test 1 | Direct connection | 99.608 us | 10.04 Kelem/s |
| Test 1 | Go proxy | 122.01 us | 8.19 Kelem/s |
| Test 1 | Tokio (32K buffer, 16 threads) | 190.49 us | 5.24 Kelem/s |
| Test 1 | Tokio (32K buffer, 1 thread) | 181.75 us | 5.50 Kelem/s |
| Test 1 | Tokio (64K buffer, 1 thread) | 136.45 us | 7.32 Kelem/s |
| Test 1 | Tokio (1M buffer, 1 thread) | 133.63 us | 7.48 Kelem/s |
| Test 1 | Tokio (8K buffer, 16 threads) | 20.219 ms | 49.45 elem/s |
| Test 1 | Tokio (w/ tokio::io::copy, 2K buffer, 16 threads) |
22.774 ms | 43.91 elem/s |
| Test 1 | Tokio (w/ tokio::io::copy, 2K buffer, 1 thread) |
25.373 ms | 39.39 elem/s |
| Test 1 | Tokio (w/ tokio::io::copy_bidirectional, 2K buffer, 1 thread) |
28.190 ms | 35.47 elem/s |
| Test 1 | std (64K buffer) | 134.46 us | 7.43 Kelem/s |
| Test 1 | std (w/ std::io::copy, 8K buffer?) |
31.662 us | 31.58 Kelem/s |
| Test 2 | Direct connection | 53.412 us | 18.72 Kelem/s |
| Test 2 | Go proxy | 80.494 us | 12.423 Kelem/s |
| Test 2 | Tokio (32K buffer, 16 threads) | 86.734 us | 11.53 Kelem/s |
| Test 2 | Tokio (32K buffer, 1 thread) | 78.640 us | 12.71 Kelem/s |
| Test 2 | Tokio (8K buffer, 16 threads) | 85.59 us | 11.68 Kelem/s |
| Test 2 | Tokio (w/ tokio::io::copy, 2K buffer, 16 threads) |
80.99 us | 12.34 Kelem/s |
| Test 2 | Tokio (w/ tokio::io::copy, 2K buffer, 1 thread) |
79.39 us | 12.59 Kelem |
| Test 2 | std (2K buffer) | 78.63 us | 12.71 Kelem/s |
| Test 2 | std (w/ std::io::copy, 8K buffer?) |
77.680 us | 12.873 Kelem/s |