Status meeting November 6, 2012¶
Dongwook: I can not longer reproduce the issue I was seeing with EM physics. Removed the unnecessary flags (openmpi, fast-math). We now have 3 EM physics implemented. They are currently ran independently ; I am working on integrating like in Geant4 proper.
Soon: I would prefer that Dongwook focus on finished all 3 models before doing the integration.
Dongwook: However I need to extend the interface before the integration can be done.
Soon: I have included Otto's code in our repository.
Soon: I have made progress in the optimization of the processes implemented. the function pow is expansive in GPU. So is a for loop with a conditional break. Just adding a for loop doing a single loop decrease performance drastically (more than factor two in my case). For fixed number of iteration the GPU is good. So this is a serious bottle-neck and is repeated more than once in the code (flow) ... and is in part due to the adaptive step size. With the current work flow this is unavoidable.
Soon: so my idea is to do just one stepping on the GPU and thus the byproduct of the GPU is the trajectory and its error and this is used to decide whether this advance or not. and return to the CPU to decide whether to increase the step size or not.
Marc: Well should we focus on just photon and electron for which we can implement all the physics.
Dongwook: yes, but what about the secondaries and the ion produced by the ionization process ; they are not really propagated but they are used in someway (details not known by us) at the very least in calculating energy loss.
Dongwook: I gleefully are not longer able to crash Cluck while compiling with CUDA (I am guessing the crash was provoked by the openmpi and/or fastmath flags).
Soon: I am still using fastmath and it does crash Cluck.
Soon: I have not yet been able to run any GPU code on gaea.
Marc: However, Jim and I have been successful by using an interactive jobs. It worked also in batch mode.
Soon: will try to reproduce Marc's result.