WMModel

WMModel

WMModel is still slightly experimental and is only included in this current build as an untested feature.

WMModel is designed to analyse the similarities between the point of high water mark between multiple trace files. By either strong or weak scaling you can provide the trace file of the HWM thread from multiple different runs at different scale, or problem size. WMModel will then generate an equation based on the problem size and core count, to predict the memory consumption at larger scale.

The model currently only works on two trace files.

To run the model requires the user to know the global problem size, and specify it to WMModel along with the trace files e.g.

WMModel --tracefiles WMTrace/trace-3.z WMTrace0000/trace-1.z --problemsize 1000000 1000000

The output will look somethng like this:

For C = Core Count, P = Global Problem.
 Y(C, P)=
        229509 + 34*C + 719.456*(P/C)

At this stage the model only models three types of relationship: Static (does not change between the files), Core based (scales proportionally to core count changes) and Problem based (scales proportionally to the estimated per core problem size).

There is experimental support for ghost cells, by specifying the problem and processor decompositions. Here we try to match allocations which have scaled according to the total number of cells, including ghost cells, per processor.

E.g. For a strong scaled 120x120x120 problem, from 32 to 64 cores, with a 2D decomposition of 8x4 and 8x8 respectively, comparing the HWM points from rank 0 on both traces.

WMModel --tracefiles 32/WMTrace/trace-0.z 64/WMTrace/trace-0.z --problemsize3d 120 120 120 120 120 120 --decomp2d 8 4 8 8