-----------------------------------------------------------------------
Minutes of MICAC reco framework subcommittee meeting - 17th Jan 2006
-----------------------------------------------------------------------
 
Attendees: Nathaniel, Robert, Sue, Jon, Sergei, Caius, George, Chris
Absent: Rustem, Brett, Jim

http://www.hep.ucl.ac.uk/~cbs/Talks/Reco/NewFramework/MICAC_Subcommittee_Meetings.html

Agenda:
1) ROOT team feedback - Sergei
http://home.fnal.gov/%7Eavva/minos/software/RootIO.pdf

2) TClonesArray vs STL - Sue
http://www.hep.umn.edu/%7Eschubert/recotest/

3) How do we converge on making a decision to work on this - All
------------------------------------------------------------------

1) Sergei presented Phillipe's tests of I/O with the new candidates. 
In summary, the I/O speed was as Phillipe expected and led Sergei to 
conclude that the things that give us flexibility (e.g. base class 
pointers, inheritance) slow down the I/O. The fastest data structures 
to I/O are vectors of concrete objects.
A number of possibilites for speeding up I/O were then suggested:
- Remove CandDigits and record them as concrete objects in vectors
- Rewrite the record structure to have fewer smaller objects and more
larger ones
- Decouple the data structure into two layers. A persistent layer of 
concrete objects and a second layer which constructs the handles.

The last solution is the best, but requires significant changes to the 
proposed framework.

2) Sue presented an I/O comparison using records containing digits and 
strips. She showed that reading a data structure containing a vector of 
handles was slower than reading one with a TClonesArray of objects plus 
reference arrays for rebuilding the handles. The speed up with 
TClonesArray comes about because memory can be allocated once and reused.
The rebuilding of the handles was also shown to be a small fraction of the 
total read time.

Nathaniel asked that the tests be repeated with vector<Digit*> rather than
vector<handle>.

3) Chris gave impressions from Collab meeting: People would be happy 
to work on a fully functional new framework to re-code an algorithm 
for example. However, the basics of the framework would have to be in 
place, and that would probably have to come from this group. Generally 
collaborators were concerned that the ntuples would still be available.

Sergei announced that Minerva had decided not use ROOT for it's software 
framework. Instead they have chosen GAUDI. Sergei's work on the new framework 
will therefore not be continued in his Minerva role.

Summary:
-------
- Sue will continue to perform speed tests with vectors and TClonesArrays.
- Given the results of the I/O tests with the current framework and the 
pseudo-loss of Sergei as a driving developmental force, the committee are 
asked to think about viable alternatives to the current proposal.
Furthermore, at this stage it may be reasonable to consider the 
restructuring of the candidate framework and the I/O model as more-or-less 
independent problems.