RE: Waveform record I/O interrupt. asyn

["Hill, Jeff" <johill@lanl.gov>]

Ø  The - already implemented - server-side CA plugins/filters in 3.15 can

Ø  also achieve this behavior.

 

An issue with this approach is that such plugins are subscription specific (i.e. client and event queue specific), and therefore the same payload, including site specific parameters together with the array data, are not typically shared, that is reference-counted, on multiple queues for multiple CA clients over multiple different filter types and expressions. Also, issues with reordering of events, subscription-xxx versus subscription-yyy, remain in R3.15 substantially unchanged, inherited from the legacy implementation which I authored. I also have some personal concerns about disruption of timing system synchronized record processing by application installed high-priority-side-of-the-event-queue per-subscription-event-filtering; I am concerned that this does not scale as the load due to a variable number of subscriptions introduced at run-time increases. Yes, the users might only install very low overhead filters to run at record processing priority, but I worry that such an architecture is opening up broad avenues for EPICS to become a less than robust system unless users are very diligent. I am concerned that a real time activity that is proven to be working might be more likely to be unexpectedly disrupted in the future by an increasing magnitude of the client load. Finally, my personal perspective is that my approach, initiated first and described in previous talks, is a more generalized architectural upgrade allowing site-specific parasitic parameters to pass transparently through EPICS from the device support data-originator to the data-consumer.

 

Jeff

 

From: tech-talk-bounces@aps.anl.gov [mailto:tech-talk-bounces@aps.anl.gov] On Behalf Of Ralph Lange
Sent: Saturday, February 22, 2014 2:10 AM
To: EPICS Tech Talk
Subject: Re: Waveform record I/O interrupt. asyn

 

On Sat Feb 22 2014 02:25:30 GMT+0100 (CET), Hill, Jeff <johill@lanl.gov> wrote:

Ø  the event queue does not buffer arrays – it only queues up the pointer to the

Ø  array. If you are reusing the same buffer (as the waveform record does) it may

Ø  also be overwritten if the client is not taking the data fast enough.

 

Ø  We are working on ways to better support very large arrays in EPICS V4 – these

Ø  must be done for us over the next 12 months – but only pieces are done now.

 

We are deploying a fix for this issue at LANSCE which is, in contrast, implemented into V3.

The - already implemented - server-side CA plugins/filters in 3.15 can also achieve this behavior.

~Ralph