Status meeting April 27, 2012

Looking at replacement of RK.

After looking at several other possible algorithm, it looks like that we do need an algo with adaptive step length. RK is a good one of course, but also Burlirsch-Stoer (and one other).

So far I looked at momentum as a way to classify the particle so that they requires the same number of internal steps. I have studied pions and will look into elections.

I am stepping back and trying to quantify/understand what we are looking for. Looking at step length vs momentum distribution. In our example a lot of the step length are around 1cm.

Concurrent kernel/Stream.

The maximum gain from using Stream is a factor 2 (maybe up to 2.5, not clear). Proving this has proven elusive because the volatility of the upload/downtime is in the same order of magnitude as the expected gain.

To get the theoritical gain, we would need to do twice has much work as the single RK steps. Estimating from the callgrind and igprof profiling, it looks like there is only 40% more work to be gained without any geometry.