V4 2nd Design Meeting Notes

2nd Design Meeting (APS) - Notes

Who is the customer?

As there are no big machines coming up and lots of existing installations, V4 will be targeting the whole community.

Compatibility and conversion issues

V4 will not support 68k processors under Tornado2, only under RTEMS. (There might be memory footprint issues, though.)

Lower limit: 8MB/68k or 16MB/coldfire - i.e. must support small and embedded systems.

Gateways will be used to convert CA connections between V3 and V4 within one system.

We should provide a set of asara records (asara = as similar as reasonably achievable) providing V3 behaviour that can be used for automatic conversion of existing V3 databases.

DBD syntax and generation

Building blocks

It seems wrong to have real C++ declarations within the dbd file. The interface definitions should rather reside in a separate header file. (That way it would be easier to e.g. have the interfaces in different languages.)

Declarations of fields and views

The V3 prompt declaration will be replaced by a "hint" or "help" text that can be applied to any entity defined in the dbd language (not just fields). Tools may use that text to display bubble help or similar things.

The V3 group declaration will be replaced by views that can be used by the configuration tool to group things into hierarchies.

Hierarchies

Jeff envisions a "global" hierarchical namespace, where a request for quad4.ps2.current.analogValue.val would be answered by a "virtual" record that forwards and collects things from different records. Such a virtual could reside on a different node. (I would call it part of middleware functionality.) It is unclear who would be writing such a record.

Agreement on: Everything before the first dot is a record name.

V3 compatibility records will define views like

view (HIHI) {
    property (....)
}

so that clients can still use the old syntax of <recordName>.<fieldName> to access fields.

Programmatic views will accept arguments (in parantheses). I.e. the programmatic "field" view will take the field name as an argument, as in

xxx.field(displayLimit)

Subproperties will be addressed using multiple dots, as in

xxx.value.displayLimits.upper

(This might map to a numeric property address - using an array of propertyIds - when transferred over CA.)

Clients should be able to find out if a certain property implements "analog", "enumerated", or "string" behaviour. (To make handling unknown properties easier .) Making the IOC answer these kind of questions induces a number of problems (versioning etc.) It might be possible to implement these "types" of properties as a conveniency layer library on the client side.

Record instance syntax

i18n

LOCALE techniques will be supported on systems that support it, on a level as close to the user as possible. (We're trying not to preclude someone from using LOCALE.) I.e. the dbd file (being a kind of source file) will need the "." character as a decimal point. It is the task of tools (like vdct) to convert any data in LOCALE format (e.g. using "," as decimal point character) to the standard format in the dbd file.

Data Access

Type info

The presented first version of a third set of traverse and find methods to find out the native types of properties does not seem to be efficient and desirable.

CA/DA Convenience layer / user interface

There will be a higher level interface library on top of Channel Access and Data Access that hides the inherent complexity and provides the end user with a simplified way to get to the structures, type, and data.

Collecting the requirements and defining the interfaces of that convenience layer is a task with high importance and priority. (Doug and Kay did a first step in that direction.)

Channel groups (with synchronized read and write operations) should be implemented as a separate class that calls the methods of the Channel class internally.

The V4 Data Store might implement an interface different from the opaque creation/deletion interface. That interface could have iterators etc. pp. to serve as a way different from dataAccess to analyze catalogs.

CA issues

Sending full catalogs vs. changed properties only

It is still unclear, if for a client's subscription all properties (i.e. the complete catalog) or only the changed properties should be sent.

Sending only changed events would certainly be a lot more efficient on the network. (Implementation idea: Bitmask at the beginning of a message that indicates which properties will be present in this buffer.) That might complicate the event system on the IOC, though. The database would have to provide a similar bitmask to the event manager, so that the CA server is able to decide what to put into the network buffer.

Sending the complete catalog might be less complicated in the CA server and clients, but the user would have no way to find out which of the properties in the presented catalog has changed.

Network protocol

Jeff will provide a description of his plans for the CA protocol.

Name server

LDAP seems to have some properties that makes it a very interesting candidate. The measured resolve rate of ~1500 per sec. sure makes people feel uncomfortable. If we can achieve >10000, it would stop this discussion. More tests with different server hardware and other DB variations (under the LDAP server) should show if there's a reasonable way to achieve the desired performance.

The way to go is to allow for other name services and implement a LDAP plug-in as a first attempt. Regular CA name resolution should still be the default and fallback.