On Fri, 2011-10-28 at 13:19 +0300, Arun Raghavan wrote: > On Tue, 2011-10-25 at 09:06 -0700, Dylan Reid wrote: > > a profiling run didn't show any hot spots. > > A quick profiling run with 5ms minreq and 10ms tlength doesn't actually > consume *too* much CPU here (~12% at 800 MHz). There might be ways to > reduce this though. > > Just from profiling, I noticed that ~30% of that time goes into ppoll() > in the kernel. I seem to recall that someone was doing patches to move > us to use epoll(), which /should/ be faster. Was this actually done? > Perhaps by the Nokia folks? If not, I'll take a look at doing this. There might have been some patches written a long time ago (over 2 years), but I think there were some problems with that, and I believe the benefits of epoll() aren't anymore as significant as they were in the past, so that effort has not been continued at Nokia. So, feel free to tackle this yourself. -- Tanu
