1. Use of Root I/O Trees for CMS Crossings
Benefits and deficiencies of Root I/O trees when : - NOT dealing with TObjects, - reading the trees entries NOT sequentially, - processing them NOT one by one.
ROOT ' 2002
David Chamont (CMS - LLR)

2. David Chamont (CMS - LLR)
Outline
Goal & scope.
Main use-case.
Crossing data model.
Persistent objects managers.
Four persistency strategies.
Storage of not-TObjects.
Implementation issues.
Performances.
Conclusions.
Wish-list to the Root team.
Future work.
ROOT ' 2002
David Chamont (CMS - LLR)

3. David Chamont (CMS - LLR)
Goal & scope
Evaluate the use of TTree for the persistency of CMS event data (whose classes heavily rely on templates and external packages).
Focus on the generation of crossings (pile-up of about 200 simulated events chosen randomly).
Not covered yet : meta-data and references.
ROOT ' 2002
David Chamont (CMS - LLR)

4. David Chamont (CMS - LLR)
Main Use-Case
Digitizer
signal event (hits)
minbias event (hits)
digis
Pom
Pom
Pom
files of ~ 500 events
events of ~ in memory.
crossing of ~ 200 minbias events.
minbias events file
signal events file
digis file
minbias events file
ROOT ' 2002
David Chamont (CMS - LLR)

5. David Chamont (CMS - LLR)
Crossing Data Model
The folders //root/pool/* represent the events composing the current crossing.
Each event folder contains collections of RtbTrackHit, RtbCaloHit,…
These collections should be kind of RtbVArray<> :
RtbCArray<> (home-made C-like array).
RtbClonesArray<> (wrap a TClonesArray).
RtVArray is an abstract class. The choice of the concrete class if delegated to the persistency manager.
ROOT ' 2002
David Chamont (CMS - LLR)

6. Persistent Objects Managers
Persistency managers are able to transfer event data (set of RtbVArray<>) between a TFolder and a TFile.
We implemented three kinds of RtbVPom :
RtbMatrixPom : for each RtbVArray<>, create a TMatrixD and attach it to a branch of a TTree.
RtbDirectPom : attach each RtbVArray<> of the folder to a branch of a TTree.
RtbKeysPom : directly store the TFolder in the TFile, each time with a different meaningful name.
I had to use old branches for RtObjArray
I was forced to split TClonesArray
RtMatrixPom is a reference, it does not use rootcint, but only root predefined classes
ROOT ' 2002
David Chamont (CMS - LLR)

7. Four persistency strategies
RtbMatrixPom (matrix)
RtbCArray and RtbKeysPom (keys)
RtbCArray and RtbDirectPom (carray)
RtbClonesArray and RtbDirectPom (clones)
ROOT ' 2002
David Chamont (CMS - LLR)

8. Storage of non-TObjects
RtbCArray
RtbCArray
I had to use old branches for RtObjArray
I was forced to split TClonesArray
RtMatrixPom is a reference, it does not use rootcint, but only root predefined classes
ROOT ' 2002
David Chamont (CMS - LLR)

9. Implementation issues
Tips to use Root together with Scram/Cms
add -p to rootcint (why not the default ?)
ask to parse namespaces (why not the default ?)
always add -fPIC for link step
remove -ansi –pedantic
Typical problems with Root I/O :
collections sizes and operator[],
storage of empty collections,
ownership of objects attached to the branches,
tuning of chain branchs.
ROOT ' 2002
David Chamont (CMS - LLR)

10. 200 random events read time (s)
Performances
Matrix
Keys
CArray
Clones
500 events file size (Mb)
119.6
63.3
118.9
63.5
119.3
63.8
106.6
57.7
500 events write time (s)
153 74
191 95
197 96
147 72
200 random events read time (s)
4.57
6.93
5.82
5.88
files are smaller than objy ones
objarray have old branches and cannot have empty arrays
clones cannot be unsplit, and there is a memory leak.
strategy 1 is interesting because it does not require the use of rootcint (but we must implement the serialization ourselves).
strategy 2 is interesting because we do not pay the price for TObjects, yet we do not take benefit from the annouced optimization of TTree for TClonesArray ( but is this optimization real ?).
strategy 3 is the most ROOT oriented, but the wrapping of objects could be counter-productive. in fact, wrapping can only rely on templates, and this currently prevent us from bypassing the streamer
An interesting result is that the 3 strategies have quite similar results…
I should evaluate cpu time without compression of data.
Les objets de base ne sont pas splites
Les collections sont precrees au sein du pool
? si j’implemente la creation de collections de facon externes a la piscine et leur
placement dans la piscine par l’utilisateur, je dois me demander ce qui se passe si
une des branches a son pointeur vide (je pourrais essayer avant chaque fill
de desactiver les branches des pointeurs vides, mais est ce que je conserve les
bons rangs ; ou alors il me faut faire au vol une collection vide (mais j’ai deja des
problemes avec les collections vides ! pour le moment, je peux faire l’hypothese
qu’il n’y a jamais de vide et toujours le meme nombre de collections.
? est-ce que les TClonesArray vides posent les memes problemes que les RtObjArray vides ?
ROOT ' 2002
David Chamont (CMS - LLR)

11. David Chamont (CMS - LLR)
Conclusions
We succeeded to read pseudo-random entries from a chain and dispatch them to few hundred folders, yet TChain/TTree/TBranch can be improved to support this use-case more efficiently.
Use of specialized Root classes (TTree & TClonesArray) improve performance but increase problems with complex C++ structures and foreign classes.
Best argument for TTree is the privileged connection to Root analysis features.
strategy 1 is interesting because it does not require the use of rootcint (but we must implement the serialization ourselves).
strategy 2 is interesting because we do not pay the price for TObjects, yet we do not take benefit from the annouced optimization of TTree for TClonesArray ( but is this optimization real ?).
strategy 3 is the most ROOT oriented, but the wrapping of objects could be counter-productive. in fact, wrapping can only rely on templates, and this currently prevent us from bypassing the streamer
An interesting result is that the 3 strategies have quite similar results…
I should evaluate cpu time without compression of data.
ROOT ' 2002
David Chamont (CMS - LLR)

12. David Chamont (CMS - LLR)
Wish-list to Root team
Enable to change the value of a pointer attached to a TBranch.
Enable to add a user-defined function to a TChain, to be called each time the current tree is changed.
Clarify/document the subset of C++ which is supported by Root I/O, and which subset is supported within trees.
ROOT ' 2002
David Chamont (CMS - LLR)

13. David Chamont (CMS - LLR)
Future work
Provide root team with unscramed demonstrations of what we observed.
Implement a fifth strategy using TObjArray, for comparison with TClonesArray.
Try to make the data classes directly inherit from TObject and compare performances with RtbObj<>.
Look inside Root classes implementations, optimize strategies accordingly, and upgrade the conclusions.
Add references between objects.
ROOT ' 2002
David Chamont (CMS - LLR)