Re: Sequencer seem to skip states with EPICS 3.16.1

[Ralph Lange <ralph.lange@gmx.de>]

On Wed, Jan 31, 2018 at 7:18 PM, Benjamin Franksen <benjamin.franksen@helmholtz-berlin.de> wrote:

> Some more testing showed that the same IOC build for linux-x86_64 is
> running as expect on my machine.
> So there seem to be a problem running this on linux-ppc.

Interesting. That confirms my suspicion that this is a problem in base,
not the sequencer; my guess is that any fast enough sequence of ca_put
operations on the same PV exhibits this problem (on this architecture).

Just a blind guess: could it be something about fast mutexes / spin
locks and how these are implemented for linux-ppc? I am not using these
in the sequencer, but they are used inside EPICS base (they have been
introduced in 3.15 or 3.16).

Another guess: wasn't there a problem report regarding put_notify a
while ago? Perhaps this is related.

Do we have tests in base for CA clients? If so, it might be possible to
adapt such a test so that it rapidly issues ca_put calls (with callback)
to the same PV and then see if camonitor sees all the events it should see.

Another angle that could be involved: does the sequencer program run on
the same IOC that serves the PV or a different one?

(Sorry for the brainstorming)

More brainstorming:

What's the setting for the thread scheduler policy? Preemptive posix thread priority scheduling makes IOCs act different than under the default Linux scheduler.

How many CPU cores do your linux-ppc and linux-x86_64 machines have?

Both could affect a low priority CA thread to get / not get scheduled while the state machine thread is switching through states.

Cheers,
~Ralph