[fabianpk]

As I understood it from the Zagreb meeting, there's some interest in how solvers perform in parallel. Since I did some limited testing on our cluster I thought I'd post the speedup rates here. First a description of the cluster:

252 nodes of
4 Xeon 5160 (Woodcrest) 3 GHz cores (dual dual core)
MPI over gigbit etehernt

The case:
64k cells, hex-mesh, about 10x10x10 cm^3 size
Conditions are above autoignition conditions, so it starts to react rather quickly.
Chemistry mechanism of 250+ reactions, 60+ species,
injection speed is 700 m/s
Courant number of 0.1, nCorrectors=3, residuals at 1e-6, limitedLinear for most div-terms, Euler in time


I've run it for 1 hour on 4 processors (which is only on one node and doesn't require network except for file access), 8 and 12 processors.

So results normalised to the 4 processor case:
processors: 4 8 12
speedup: 4 7.3404 10.5819

Come to think of it, this isn't really a test of the lagrangian particles in parallel, as it's only max 50 parcels, and they are most likely always on just one processor (the one closest to the injector). I used a metis decomposition.

If I have time I'll run some of any of the tutorials but with a finer mesh, so it's easier to compare.