Thank you for warm tone of your reply and flattering assumptions about my knowledge.

Actually, I have written some kind of proxy, which uses AE::HTTP for backend querying. There I've met the limitation HTTP::MAX_PER_HOST = 4

You request can't be cancelled inside callback, because it is invoked only after http call (successful or unsuccessful).

I think, I do understand well what is load testing and how AE works. It works in single process CPU thread, that means, that it gets some performance boost only during I/O wait operations via processing or enqueueing other requests. So, might reach the limit of parallel requests much earlier, then you'll get 1000 simultaneously running http requests, due to limitations of the single core CPU. So, perl / AE scheduler might be just overloaded in doing http requests, processing results, and switching between contexts.

It may be better to gradually rise the number of simultaneous http requests. According to that metrics: https://github.com/Mons/AnyEvent-HTTP-Server-II AE::HTTPD (written in similar manner as AE::HTTP), serves only about 560 rps (on single CPU core, as I understand). That explains well, that doing 1000 dummy requests, will load your CPU to 100%.

I can advice you to use ab from apache-tools package for simple load testing, because it is written in C, and aimed for loading.