Re: Experience with CAJ

[Kay-Uwe Kasemir <kasemirk@ornl.gov>]

On Sep 25, 2006, at 13:43 , Christopher Larrieu wrote:

> Does anyone have experience using Cosylab's java channel access  
> (CAJ) in a
> real-world, production system?  I'm curious about using it in lieu  
> of JCA.

Hi Chris:

Are you the same Chris who used to be on the 'other' coast of the US?

At the SNS, we had problems with the pure java CA client lib when
it comes to IOCs that use CA security.
The CA server on the IOCs would enter a state where it's hung.
New CA connections would create new client threads, but each new
thread hangs waiting for some semaphore inside the access security
library's asAddClient() routine.
Try searching previous tech talk messages:
http://www.aps.anl.gov/epics/tech-talk/2006/msg00503.php,
and look around April 25, 2006 for messages from Jeff and me.

The symptom: No new CA connections, and a 'caget some_PV' will issue
the 'timeout' error messag after about 1 second
(as common for missing PVs), but hang for another 29 seconds before  
exiting
(symptomatic for this error where the new client thread is created  
but then hangs).

Since this is very bad for operational IOCs, the only remedy being a  
reboot,
we quit using CAJ for now.

The problem is probably caused by CAJ sending user/host information  
in the 'wrong'
order or at a 'bad' time, but with CA it's of course a bit difficult  
to determine
what's a 'correct' message or the 'right' time ;-).
At one time we caught the error with a version of EPICS base that was  
patched
to allow a 'semShow' on the access security semaphore. I think it showed
that another CA task held the semaphore, but I can't find the detail any
longer. Maybe Jeff remembers, I'm pretty sure I emailed him the results.

I also saw occasional messages on the IOC
  CAS: request from someIP:port => "bad resource ID"
which only occurred when using CAJ.

So while I think CAJ is terrific, I would only use it through a gateway
until these kinks are worked out.

With the JNI-based CA client library for Java, there are also issues
that Tom Pelaia at the SNS and Matej at Cosylab are currently  
investigating:
As far as I understand, the JNI code includes a list for the incoming  
monitors.
If the java client code doesn't digest them fast enough, that list  
grows ad infimum,
instead of engaging the CA flow control.
We first thought it's a memory leak, since the app keeps growing.
It's not, just a list that grows, but nonetheless fatal.

-Kay