Zed, would you stop with the strawman bullshit already? Nobody serious ever claimed epoll was always faster than poll. The very graphs you linked to in your o

Zed, would you stop with the strawman bullshit already? Nobody serious ever claimed epoll was always faster than poll. The very graphs you linked to in your original post, the ones that you said were "flat out" wrong, themselves showed that poll was faster with smaller numbers of inactive fds on certain benchmarks. And you don't need a benchmark to find it out, either. Just think about it. Poll doesn't scale because even when you have a ton of dead connections, you still have to iterate over every single fd that you're tracking when you're notified of an event. The array that you get back from epoll only contains active fds. As the active/total ratio goes up, you're eventually going to be doing just as much work as you would be doing with poll (iterating over every single fd), except you're also going to have taken the hit of the work to support epoll itself. In other words, when ATR reaches a certain threshold, setting aside the possibility of a confounding (there you go, I used your word du jour) implementation detail, there's no possible way epoll can be as fast or faster than poll. It will always be slower. And if you never realized it, it's simply because you didn't stop and give it 5 seconds of thought, not because you were lured in by the malicious lies of the ignorant epollers. The only interesting thing your original post illustrated vis-a-vis poll vs epoll was what the actual threshold was for your particular machine and benchmark. Superpoll is an interesting idea, but there are serious issues with the approach you described. By adding the overhead of managing the epoll/poll migrations, you're inherently going to negatively impact the pure performance of both. I might have read it wrong, but it also sounded like you intended to start connections off in poll and migrate them over to epoll as their activity ratio went down. But with that approach, if you're getting spiky traffic that hits once and then just waits for the connection to time out, you're getting the absolute worst of poll, and then possibly spending the time to migrate over to epoll where it's just going to wait to die anyways. Anyways, getting a negative response is hardly justification for posting someone's picture and ridiculing his appearance. It's that kind of behavior that makes people decide you're not worth dealing with regardless of how good your code is or how smart you are.