Hello All:
I am working on a project which required writing a driver.
The hardware is an embedded VersaLogic PC-104 system (ISA bus).
The processor is a Pentium M, 1.6 GHz with 512 MB memory.
Operating system is Slackware 12.0 Linux with a 2.6.25.7 kernel compiled with the low latency preemptible kernel.

One of the system tasks involves operating a rotary knife device to size a product moving at 1,000 Ft. / Min. ( 200 In. / Sec). Our desired interrupt period for the driver is 62.5 uSec. generated by a counter / timer on one of the ISA I/O boards.

At times, the driver's clock interrupt may not run for >1.2 milliseconds. This doesn't happen very often and when it does happen, it has to happen at the wrong time to degrade our accuracy. But as all of you know, if it can happen, given enough time, it will happen.

How does this latency compare to what some of you may have experienced?
Any suggestions for improvement? Incidentally, running without X doesn't improve things. The latencytop utility apparently only works for user-land processes, so it didn't help.

As an aside, I have written a simulator to test the system using real-time Linux. The real-time Linux system is a 1.8 GHz Athlon that is happy running a 31.25 uSec period interrupt with ~ 15 uSec. max latency - roughly a 80x improvement. I usually use real-time Linux for something like this, but this customer didn't like the idea or at least wanted me to give a standard Linux a real good try first.

Thanks
wevl

I think that this is the problem you should address. Why is the interrupt occasionally masked for 1.2 milliseconds? Is it masked in the hardware or is it masked in the kernel? If it is masked in the kernel then you need to look at what interrupt is taking 1.2 milliseconds of uninterruptable processing to clear the interrupt. Is the masking interrupt some part of your driver system or is it something completely unrelated like a system clock interrupt?

--------------------
Steve Stites

Does anyone have any experience with non real-time Linux and driver latency expectations? 1.2 to 1.3 milliseconds may be about as good as one could expect. I just don't know.

Read this thread and follow the links. Your questions will be answered, though you may not like the answers. Or, rather, your client may not like the answers.