aquatic VS wt-tracker

How does Bitmagnet compare to Aquatic? https://github.com/greatest-ape/aquatic

I have a question on UnsafeCell usage that popped up while implementing io_uring support for aquatic_udp. I find the docs slightly confusing (in particular the part that I've marked in bold):

I've done a new round of benchmarking of open UDP BitTorrent tracker implementations. Results were great for aquatic_udp, achieving double the throughput of opentracker when running on 8 CPU cores:

If you run your tracker on Linux and an info hash whitelist approach would work for your use case, it might be worthwhile having a look at aquatic_ws [0]. It relies on tungstenite [1] for websockets and achieves around 20x the throughput of the reference implementation when running with four threads.

[0] https://github.com/greatest-ape/aquatic

aquatic, a multithreaded BitTorrent tracker with world-leading performance

But aquatic_udp doesn't use glommio.. https://github.com/greatest-ape/aquatic/pull/29

aquatic is a Apache 2.0-licensed BitTorrent tracker written in Rust that I have developed over the last couple of years. It focuses on correctness, stability and high performance.

The udp request parsing code seems to have been copied from aquatic. Please note that the Apache 2.0 license requires attribution.

wt-tracker (nodejs, single-threaded) reached 31 thousand responses a second

aquatic_http is a BitTorrent over HTTP over TLS tracker and aquatic_ws is a WebTorrent over TLS tracker. I don't have full current benchmarks comparing them to other options, but in a previous one, aquatic_ws handily outperformed both wt-tracker and bittorrent-tracker.

I did some benchmarks with my load testing tool and got very encouraging results: 1.8M tracker responses per second when using 14 threads, compared to a maximum of 117k responses per second for closest competitor wt-tracker. I suspect that the openwebtorrent tracker would be a lot closer in performance to aquatic, but since it only runs over TLS, I couldn’t test it, as the load tester doesn’t support it. More info on the benchmark (PDF).