Experimental Physics and
Industrial Control System

Ru Igarashi via Tech-talk <tech-talk at aps.anl.gov> · Wed, 14 Apr 2021 10:09:42 +0000

On Wed, 14 Apr 2021 10:06:27, Ralph Lange via Tech-talk <tech-talk at aps.anl.gov> wrote:
>On Wed, 14 Apr 2021 at 09:04, Ru Igarashi via Tech-talk <tech-talk at aps.anl.gov> wrote:
>>Thanks for the -no_cache suggestion.  It did stop the problem but is not
>>a long term solution because I'm relying on caching so that the
>>gateway can act as a proxy or channel access concentrator (to
>>reduce traffic).

>The '-no_cache' option stops the Gateway from answering CA get() requests directly from
>its internal data cache. Instead, the Gateway will forward all get() requests to the IOC and
>relay the IOC answer back to the client.
>This option does not change the Gateway behavior regarding subscriptions, i.e., it will still 
>"concentrate" many identical subscriptions by its clients into a single subscription to the
>IOC and do the fanning out of updates.

>Original motivation for the '-no_cache' option was the chance of visible inconsistencies
>introduced by the cache, where - assuming an analog value with a deadband for monitor
>updates - a client would get a different value directly from the IOC (the actual current
>value) than from the Gateway (the last monitor update from the IOC that may differ by
>the deadband width).

In principle, this should not be the mechanics causing my problem, since no
deadband is in play, and the value is clearly updated in a get immediately
after the put.  It's similar in that the client does not see what is on the IOC,
but that's despite a valid update, and as it turns out, a corresponding update
does get emitted by the gateway but the data is empty/null (I was expecting
to see the client report zero rather than the old value, so I'm missing
something about how the data payload is interpreted).

I need to reiterate that this only happens for one combination of OS and EPICS
version of IOC and gateway.  All others seem to be orders of magnitude less
affected, if at all. 

>Again: For subscriptions, '-no_cache' or not should not make any difference.

My concern is not with subscriptions, but with the many bursts of batch get requests
from periodic loggers, as well as the continuous (possibly low level of) random
get requests.  What I would have to evaluate is how many of these currently
require the gateway to reconnect anyways (when there isn't an existing
subscription to piggyback off of).

>Cheers,
>~Ralph

Experimental Physics and Industrial Control System

Experimental Physics and
Industrial Control System