A short note, it seems we get pagefaults in rt context, latency-test says so.
I will build latest and run some tests. The intriguing thing is all other rt programs run fine, just linuxcnc that overshoots. / regards, Lars Segerlund. 2013/6/13 Alec Ari <[email protected]>: > Hello everyone! I just talked to Lars Segerlund over the course of a few > private emails, we've narrowed down the "Missed scheduling deadline" problem > for LinuxCNC with PREEMPT_RT to an issue within RTAPI. A lot of people (if > not everyone) has this problem. I have used debian kernels, I have compiled > nearly 100 different kernels, changing settings in regards to ACPI, the RCU, > every option that has a direct impact on latency, whether positive or > negative, has been changing, with absolutely no change in latency. > > Several kernel versions have been tried, such as the official debian > PREEMPT_RT kernel, and several custom kernels that I configured completely > from scratch (all options by default disabled, allowing me to turn on only > the ones I need to boot my system) and no change what-so-ever in regards to > pagefaults and latency. > > This is no time to say "don't use gentoo" "don't compile your own kernel" > "your kernel config is broken" etc, this is an issue that relies within the > LinuxCNC tree. I have been hacking away at RTAI and PREEMPT_RT for a few > years now to know how they both work, inside and out. I am not however, > experienced with LinuxCNC itself, but I can guarantee that this issue is not > relevant to Linux distributions, kernel releases, kernel configs, tool chain > issues, and so on and so fourth. > > After running cyclictest for hours tonight (12 threads) while in the > background compiling kernels, GMP, GCC, GNU Mpc, all with 50 parallel > compiling jobs, on top of all that, running a CPU burn-in utility doing > excessive number crunching with the use of exactly 1,000 threads, with > integers in the billions. Latency with cyclictest reached no higher than 110 > microseconds, and that's with using the standard POSIX interval timer, and > not even a fully tweaked kernel. > > The latency-test program in LinuxCNC almost instantly shoots up to over > 200,000 nanoseconds upon interacting with the system at all (idling it can > last a bit longer.) Launching LinuxCNC itself throws several errors as well, > "unexpected real-time delay, check dmesg for details" (which for the record > doesn't mention anything..) and basically says the same thing in the terminal > window, just worded a bit differently. > > Lars Segerlund and I are going to be tackling this issue the best we can, but > any changes to do with LinuxCNC are going to have to be by people besides > myself. Changes that have to do with the kernel side of things, that part I > can do. Any additional help is greatly appreciated! > > Alec Ari > > ------------------------------------------------------------------------------ > This SF.net email is sponsored by Windows: > > Build for Windows Store. > > http://p.sf.net/sfu/windows-dev2dev > _______________________________________________ > Emc-developers mailing list > [email protected] > https://lists.sourceforge.net/lists/listinfo/emc-developers ------------------------------------------------------------------------------ This SF.net email is sponsored by Windows: Build for Windows Store. http://p.sf.net/sfu/windows-dev2dev _______________________________________________ Emc-developers mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/emc-developers
