Hi On Mon, Apr 20, 2009 at 2:52 AM, Lennart Poettering <lennart at poettering.net> wrote: > > For example, Nokia recently found out that one of the biggest > bottlenecks (or I guess literally, smallest bottlenecks ;-)) on ARM > processors is actually PA's use of poll(). Replacing that by epoll() > already save quite a bit CPU time. > In fact, it all depends on the usage of PulseAudio, and your target device. In the case of maemo and poll(), it is because we have a high number of client contexts, and a low latency setup for voip cases. If you combine both, you can easily get 600Hz checks of 30 fds. That, indeed, can take some CPU time.. FWIW, the epoll() patch is not ready for upstream yet. regards, -- Marc-Andr? Lureau