participants: R.T. Jones
started February 20, 2005
last updated February 24, 2005
| A quick Geant3 vs Geant4 speed comparison for 10 GeV electrons (800MHz pentium III, 256MB) | |||||||||||||
| |||||||||||||
A comparison between Geant3 and Geant4 has been carried out by BaBar [2] which showed excellent agreement between the results produced using the two simulation packages. BaBar did not give performance figures, but indicated that the two packages have comparable speeds. A more recent quantitative comparison is provided by the CMS simulations group [3] for tracking charged muons in the magnetic field. Their results are given in table II.
| Time (msec/event) to track 20 GeV mu+ through detector, same number of steps in Geant3 and Geant4 (2GHz Athlon cpu) | |||||||||||||
| |||||||||||||
Regardless of relative performance, it is the enhanced capabilities of a simulation tool under active development within the NP/HEP community that will ultimately constrain our choice for GlueX simulations. In terms of enhancements present (or absent) in Geant4, the following stand out as being of potential interest to GlueX.
One potentially interesting project that has been spun off from the Geant4 development is the GDML (Geometry Definition Markup Language) project [4]. This project is based on the same idea as the GlueX XML geometry package hdds but has been created from scratch by the Geant4 development group at CERN. The plus side of this is that the GDML project provides a tool to convert a detector geometry written in GDML directly into the c++ classes needed to express the geometry in Geant4 and ROOT. This parallels exactly what has been created for Geant3 and ROOT within the GlueX hdds package. The down side of this would be that adopting GDML would require an extensive re-writing of the GlueX XML geometry definition from hdds to GDML.
Before tackling this potentially time-consuming task, I decided to check out what other experiments are saying about GDML. A quick web search turned an interesting report [5] from the STAR experiment at BNL that tackles this exact question by presenting a detailed comparison between GDML and AGDD (Atlas Geometry Definition Description). The AGDD was developed in the late 1990's and is the basis that was used for hdds. In fact, hdds is just AGDD with some extensions for GlueX, so the two are virtually the same thing. The jist oft the STAR report is that GDML seems to have been created by the Geant4 group in a vacuum, and has not yet seen much community use. A couple of notable features of AGDD (that are incidentally very useful to GlueX) are missing in GDML, and the report argues that the new features in GDML are well covered by existing features of AGDD. They conclude that GDML is inferior to AGDD as presently conceived, and substantial involvement from experimental groups will be required to turn it into something with real practical utility. It is gratifying to see our early choices in GlueX software being borne out, but also that mainstream HEP developers are pushing ahead with XML geometry projects. If we press ahead with our own transition to Geant4 we might make a real contribution to the GDML project. Alternatively we might benefit from taking it slow and waiting for mainstream developments along these lines to mature.