On-Site Rust Channel Benchmarking Helper

Deploy on server to determine which public crates are the fastest for communicating in different architectures

Requirements:

• 2 CPUs are needed for most of the experiments. Please refer to main() to change which cores you'd like to be utilized before running.

Analysis Pipeline

1. cargo run --release
2. Once the experiments are completed, run python3 analyze.py to view the results and the corresponding latency distributions

Crates tested by Architecture:

Async

• Kanal (https://crates.io/crates/kanal)
• Flume (https://crates.io/crates/flume)
• Tokio (https://crates.io/crates/tokio)
• Async_Channel (https://crates.io/crates/async-channel)
• Crossfire (https://crates.io/crates/crossfire)
• Thingbuf (https://crates.io/crates/thingbuf)
• Tachyonix (https://crates.io/crates/tachyonix)
• Postage (https://crates.io/crates/postage)
• Async_Broadcast (https://crates.io/crates/async-broadcast)

Busy-Spinning

• Rtrb (https://crates.io/crates/rtrb)
• Ringbuf (https://crates.io/crates/ringbuf)
• Crossbeam-channel (https://crates.io/crates/crossbeam-channel)
• Crossbeam-queue (https://crates.io/crates/crossbeam-queue)
• Kanal (https://crates.io/crates/kanal)
• Flume (https://crates.io/crates/flume)
• Omango (https://crates.io/crates/omango)
• Npnc (https://crates.io/crates/npnc)
• Magnetic (https://crates.io/crates/magnetic)
• Tokio (https://crates.io/crates/tokio)

Future Plans

I'd like to split up the experiments to test mpmc / mpsc / spsc against each other in these different architectures. This will be included at some point but for now mpmc & mpsc have been favored as there are more use cases for them

Notes

• Some of these crates (such as async_broadcast) are tailored to specific features which incur some speed disadvantages. Thus, I recommend doing adequate research to determine the best crate for your architecture and only use this as a reference.
• Thread-pinning has been included in most of these experiments so please check the main() function and change the parameters as needed. If you'd like more realistic non-jittery results, consider using isolcpus or the like
• payload_size parameter was included in main() as well, feel free to change this at your discretion. The default is 0 bytes

P.s.

If you'd like another crate added, have questions / ideas, please DM me on twitter @Dub0x3A or on discord 0xdub. Cheers and thanks