timvisee / pixelpwnr-render Goto Github PK

Right now you're spawning a thread for each cpu core (with thread::spawn(...) and executing a worker on it. Then for each worker you create a tokio cpu_pool with 8 (hardcoded) threads. This means that you will, on an 8 "core" system, spawn 64 threads, meaning a lot of context switching.

The tokio cpu_pool is actually designed to scale Async I/O futures across available CPU cores. Removing the second layer of thread pooling decreases memory consumption by about half and increases throughput (on my laptop, binding to 127.0.0.1 by about 30%