I'd avoid the cluster module, its not the recommended way to scale Node.js. It exists mostly as a way to make naive Node.js benchmarks perform better in multi-CPU comparison testing... like yours! :-) Many cloud providers charge per-CPU, so node's "mostly not threaded approach" is reasonable, IMO and that of many other users. Node is scaled by spinning up more instances, each instance being one of the cheapest "one CPU" variety. I'd be more interested in seeing a version of your benchmark that limited each of the language instances to a single CPU, and/or that spun up multiple node instances equivalent to one multi-cpu instance of go/elixir --- this latter may sound weird, but its a "cost equivalent" comparison, which is ultimately what's important: transactions served per $$.

FastHTTP is the fastest server for Go, it's going to crush Elixir by something like 10x, it's really fast like the fastest c++ / rust libraries.

Check out https://github.com/valyala/fasthttp

Well I believe few people use the built in http module in go. Not sure if you're testing would allow for 3rd party frameworks but the Iris framework loves to claim fastest go web server. Gorilla or gin are also popular.

As for the structure of it you would like have everything split out into goroutines with a worker pool of goroutines ready to ferry the data from request to backend and back to client.