<I'm backtracking a bit> On 1/31/2012 10:49 AM, gene heskett wrote: > On Tuesday, January 31, 2012 09:53:43 AM andy pugh did opine: > >> > On 31 January 2012 16:36, gene heskett<ghesk...@wdtv.com> wrote: >>> > > In grub, if the rtai kernel line has "isolcpus=1" appended, which >>> > > takes cpu1 out of the scheduler, then after the boot in completed, >>> > > everything is running on cpu0. >>> > > >>> > > Then, using taskset, put emc/linuxcnc to running on the now forcibly >>> > > idled cpu1. >> > >> > I thought LinuxCNC did this automatically if isolcpus was set? > Apparently it does not, at least not on my box, Andy. There is a line in > the dmsg output that seems to indicate it is aware of isolcpus, but IMO its > wrong. In fact I'll look and see if taskset's use fixes that right now. > No, that line: > > RTAI[hal]: mounted (IPIPE-NOTHREADS, IMMEDIATE (INTERNAL IRQs DISPATCHED), > ISOL_CPUS_MASK: 0).
Gene: I haven't spent enough time in the RTAI source code to know just what ISOL_CPUS_MASK means and what it causes to happen or not happen. Perhaps this will turn out to be a vital clue but I don't know how to interpret it. As for Andy's reply, what LinuxCNC does automatically is start up the *realtime* processes on the isolated cpu1, not all of LinuxCNC. Many processes are started on cpu0 as you can see using, e.g., htop. These might include axis, milltask, etc. On my ASUS board right now, running 2.6.32-122-rtai kernel with isolcpus=1, I started htop. I started the do-nothing, loop-forever "cpu hog", which drove the cpu0 usage to 100.0% and keeps it there. I started LinuxCNC 2.4.6 with the 3-axis mill simulation. I started Firefox for good measure. According to htop, I'm consuming about 430MB out of 2003MB of RAM, while the cpu hog accounts for roughly 50% of cpu usage, axis roughly 25% and everything else is less. cpu1 usage, of course, is registering 0.0% (except for a sub-1% perturbation once in a while). (Incidently, I also see the "ISOL_CPUS_MASK: 0" message). Response to the PS/2 mouse and keyboard remains satisfactory. I can move windows, surf the internet, click buttons in axis, etc. No problem. If I kill LinuxCNC and invoke the latency-test instead, memory usage drops about 30MB while cpu0 usage remains at 100.0% as expected. Now the lion's share is consumed by the cpu hog because there's nothing to compete with it. The reported latency jitter numbers are excellent, slightly less than 2000 nanoseconds. I'm thinking about your remark that the behavior (on your system) of the two pre-release builds has changed recently. Did this change of behavior occur after you brought up your two-way NFS services? I'm wondering if you're sacrificing cycles to an angry NFS god. What does htop say? I'll try later to pull down a pre-release 2.6.0 build and see if that changes the picture for me, but so far I haven't been able to replicate the isolcpus-related portion of your original problem statement. The probing problem is a different kettle of fish. Fortunately you've got really good folk looking after that one. Regards, Kent ------------------------------------------------------------------------------ Keep Your Developer Skills Current with LearnDevNow! The most comprehensive online learning library for Microsoft developers is just $99.99! Visual Studio, SharePoint, SQL - plus HTML5, CSS3, MVC3, Metro Style Apps, more. Free future releases when you subscribe now! http://p.sf.net/sfu/learndevnow-d2d _______________________________________________ Emc-users mailing list Emc-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/emc-users
