Ok, technically you are right, that is good answer to my question.
Usually in my case messages are small and "rare".
i.e. there is on master process, which receives messages from child processes. And select() loop in master process. But when master determines that there is the message from some child, it tries to read full message, without trying to multiplex between two childs and reading message from both at time.
So, after select() told filehandle is readable, I read whole message from child, and during that process do not select() other handles. Child process (tries) to write whole message at once and does not do heavy job in between. It's also a pipe, i.e. on localhost.
Even if there is 100 children, and they do actual heavy job, and send 100-bytes message to master once a second, it's just 100 messages perl second.
That approach of course, is not good when you're writing thing like http proxy (like nginx) with slow clients, when multiplexing between clients directly affects performance.
So for things like http proxies (like nginx) there should be more complicated solution.