Hello, This week I have tested my CPU topology detection on different platforms and I found some bugs and fix them on AMD (thanks to profmakx from IRC for giving me access to his 8-core AMD bulldozer). I also implemented a schedule policy as following: - when a process is scheduled, it first looks for free CPUs, which have their sibling free too (both threads are free) - if not found, then it searches for the process who has the lowest lwp_batch (so a process which is not so CPU hungry) - added a sysctl option to enable/disable the HT awareness of the schedule. Also a little refactoring: created a sysctl branch usched_bsd4 where I added all usched_bsd4 specific parameters.
Those two rules stabilized the time needed to run when there are only two processes (they are scheduled on different cores). I will come tomorrow in the afternoon with some tests. I will continue testing my implementations on vkernels and compiling without SMP support (I found that there are some compile problems). Thanks, Mihai C.
