kt315_

While no 503s were observed in the test, there were timeouts. I updated the article with corresponding graph.

None of the tested webservers were returning 503.

Sleeping is done with setTimeout:

https://gitlab.com/stressgrid/dummies/blob/b342b02407ce09cec...

This isn't busy wait. Instead it yields for specified time period, much like network request to the backend database would do.

We are preparing new benchmark test for major platforms. Among other suggestions it will include memory consumption.

Author here. We tried number of acceptors that was 4x and 16x number of cores without any difference.

With Gun it is also possible to create a specialized process structure albeit with more heavyweight processes behind each connection. Mint provides lightweight (possibly more efficient?) abstraction to achieve similar results. Thanks Eric and Andrea!

I think Gun has very similar philosophy, e.g. it is connection oriented and does not impose pooling and such.

What is the advantage over Gun? (https://ninenines.eu/docs/en/gun/1.3/manual/)

One of the benchmark goals was to test "scheduling" efficiency of each runtime. In other words to show how well it scales given many-core instance, which often is more economical in transaction/$$ sense.

Question: how using cluster module is different from spinning up multiple node instances?

What would be the good Java web server to test?

Good point, will add.

Author here. Planning to run the same test using cluster module with one worker per CPU.

What would be the most performant way to serve HTTP in Go?