Project

General

Profile

Meeting

Date: June 06, 2014 at 9:15 AM US CST (Fermilab local time)
Place: 1-866-740-1260 (ReadyTalk line) (Host : Soon)

Participants:

Andrea Dotti
Krzysztof Genser
Soon Yung Jun
Boyana Norris

Discussion

(Summary by Soon & Krzysztof)

We discussed Boyana's new results related to the introduction of a
helper class G4xyd in G4PhysicsVector and comparing it to Geant4 v10.0.

We mainly concentrated on the data in the context of the the function
affected by the change the most, namely
G4VEmProcess::PostStepGetPhysicalInteractionLenght. It looks like TLB
and L2 data/instruction miss rates get worse while the L1 data cache miss
rate has improved due to the change.

Overall, the results show usually a small ~0.5% improvements in the
G4PhysicsVector::Value metrics (where we would expect most of the effect)
and at best mixed results for other functions, again at the ~0.5% level.
For comparison, in Geant4 v9.6.r07 case, the result was ~1.5% time improvement.

The native hardware counters (in addition to the Performance
Application Programming Interface (PAPI) counters used here) may help
to understand the observed (mixed) performance results.

For the next meeting Boyana will try to provide timing related
information to complement the above result.

(note that the function name
G4VEmProcess::PostStepGetPhysicalInteractionLenght is truncated to
G4VEmProcess::PostStepGetPhysicalInteraction in the presented tables)