Ok, not a kernel bug! My last question 'default taskset is set to mask 8, is that ok?' is the key. I have a 4 cores system and mask 8 means the 4th core (the one beeing actually used).
I changed all the proccesses affinity to 'f' and then the processes started to spread in all 4 cores, as it should. Funniest part is, default_smp_affinity is set to 'f', so I don't know why proccesses start with affinity 8 2012/4/30 Wido <[email protected]> > Hi again > > 2012/4/29 Ben Hutchings <[email protected]> > >> On Sun, Apr 29, 2012 at 01:44:56PM -0300, Wido wrote: >> > I tend to follow new kernel feature from 'the H online' and I don't >> recall >> > seeing something like this was ever mentioned. I know current default >> sched >> > is the BFS, but it doesn't mention such behavior. >> >> The scheduler is CFS. BFS is someone else's project outside of >> mainline Linux. > > Then I misunderstood what I read. I know it's from Con Kolivas, I just > though it is set as default. My bad :D > > >> > A little test I did was opening a youtube video (I use chrome), the >> cpu was >> > almost 100% but, in htop. around 70% was show as low pri. If I run >> another >> > program, lets say VirtualBox, they start to share the same core, >> instead of >> > using another one that is free. >> >> That does sound wrong. However it is possible that the CPU frequency >> is being changed so that that single core still has enough cycles to >> run both programs without slowing them down. > > My CPU has the ability to powerscale, but almost always is running at max > speed. I have a gadget to play with cpufreq =) > > >> > This behavior seems buggy to me. I've been using Debian since Woody, >> this >> > desktop has almost 1 and a half year, and this started to happen when I >> > installed 3.1 >> >> Here's how you can test this: >> >> 1. Start the video playing >> 2. Measure how long it takes to boot a particular VM in VirtualBox >> 3. Repeat this another 4 times, so you have a total of 5 runs >> 4. Use taskset to force the video player and VirtualBox onto separate >> cores >> 5. Repeat another 5 times >> >> By comparing the two sets of 5 times, we can see whether the scheduler >> should be using more than one core. (Also, if the video plays more >> smoothly in step 5, then the scheduler should be using two cores. >> But this is not as easy to measure objectively.) >> >> You can substitute whatever else you're interested in for steps 1 and >> 2, just as long as it is something repeatable. > > I didn't need to run 5 times. The processes tends to bond to core 4, I was > playing a flash game and copying files from another pc (using samba) and > when I moved chrome and dolphin to cores 1 and 2, respectively, they > started working a LOT better, from almost non responsive to really smooth. > > SAR is not showing much yet, will leave the PC on and will run some more > things (with and without tasksel) to show differences on the cores usage. > But apart from that, not sure what other info can I send. > > > > Ben. >> >> -- >> Ben Hutchings >> We get into the habit of living before acquiring the habit of thinking. >> - Albert >> Camus >> > > BTW, default taskset is set to mask 8, is that ok? > > cheers > -- > Wido > -- Wido
