Present:
Marc, Giulio, Lorenzo, Pere, Xin, Kate, V.
Marc introduced his work.
Followed
discussion about sharing memory between processes.
Marc is working on a factory model.
Pere propose some sort of wrapping malloc or default allocator in a transaction model.
Marc will present his prototype at next meeting (in two weeks)
Giulio described his work in CMS.
Use perfctr on AMD.
main finding:
1) CMS code: a lot of code, trash of instruction cache.
2) speed up factor 10 since last year: 10 times less instruction, processor used worse...
current activities
Big library:
report in two weeks
different malloc
20% improvements
lower priorities in CMS
Lorenzo reported on minuit (already presented at the OpenLab Multi-thread workshop end of may:
MPI version developed by BaBar
Lorenzo migrating to OpenMP
works going on also on Genetic-algos
Pere is working on multiprocessing in Python:
two modules exits processing and parallel-pyhon
He will circulate a talk he has prepared for LHCb
Xin introduced his new work on multi-threaded G4.
Main problem in G4 is that the same class contains both long-living components common to all events and and short-living components that shall be assigned to each single thread.
More next meeting (Gene will be visiting CERN at that time
Kate introduced her plan of work of using perfmon to reproduce Lassi/Giulio performance measurements on our INTEL multicore
machine.
She plan to write a generic python script that can be then used also for other applications.
Later she plan to analyze code and start producing examples of code that best exploit multicore architecture even in single-threaded
applications.
Next meeting July 3
Actions
VI: establish a mailing-list
VI: find an indigo home
NN: compile a list of allocators and memory managers
NN: investigate the development of a tool to check the lifetime of memory blocks and how they are shared among processes
There are minutes attached to this event.
Show them.