Choose timezone
Your profile timezone:
System performance modelling WG meeting
→
Europe/Zurich
31/S-028 (CERN)
Participants
- Local: Andrea Sciabà, Markus Schulz, Davide Costanzo, Dirk Duellmann
- Remote: Frank Würthwein, David Lange, Helge Meinhard, Johannes Elmsheuser, Andrea Sartirana, Oxana Smirnova, Yves Kemp, Gareth Roy, Graeme Stewart, Costin Grigoras, Pepe Flix, Concezio Bozzi, Michel Jouvin, Alessandra Forti
Task list review
Glossary
Everybody is invited to read and comment what is there. In particular Markus suggests experiment representatives to check if some terms are defined with the proper meaning for their experiment.
Most important workloads
ATLAS (Johannes): refer to doc. A list of detailed references has been added after the meeting.
ALICE (Costin): refer to doc.
Costin confirms that they will use the Alpha messaging framework but many details have still to be defined. Concerning the data format, it should not be root-based and it should be as memory-efficient as possible.
LHCb (Concezio): refer to doc.
Packaging applications
Andrea S. reported that he could successfully build and run a container using the machinery developed last summer by Fabio Luchetti (a summer student) for the HEPiX benchmarking WG.
Workload properties
Some guidelines on what information to provide about metrics are provided in the doc. Markus suggested that somebody should play with a job and write examples on how to get these metrics.
Cost evaluation process
Jan "solved" Markus' exercise in the doc. Gareth provided some estimate for a UK Tier-2 (Glasgow).
Toy model of a workload
Markus thinks that using UML sequence diagrams would be a very effective way of describing how a job behaves. He will use an ATLAS MC DIGI+RECO job as an example and have it reviewed by somebody in ATLAS.
WLCG/HSF workshop
The workshop will give us a chance to inform people outside our WG of what we are doing. Instead of having a series of reports, Markus suggests to show a simple workload as an example and go through all the steps in the modeling. Even if simplistic, it will show in which areas we are lacking more and need to invest more effort.
This could take half of the time at our disposal, and the rest would be for discussion.
There are still some free slots in the agenda, we could use them for brainstorming activities in smaller groups (open to external contributions).
Frank stresses the importance of showing which are the parameters the affect the cost, without going into too many details. It should be easy by doing so to understand why we are falling short in resources.
Dirk comments that it would be good to have a model with few parameters that we can fit to past data to test its predictive power. For example, Markus adds that one could look at past extrapolations of processing time vs. average pileup. It might be an interesting small project to start now and to present at the workshop.
There are minutes attached to this event.
Show them.
