Example of a parallel acceleration anomaly

parallel-processing

Solution

Consider a matrix algorithm that for the purposes of parallelization is modified in such a way that every computing node is assigned a region of the matrix and only operates on this region.

If with one node the region doesn't fit into cache (because it's the whole matrix) and with multiple nodes it starts fitting into cache (the more nodes the smaller is the region) accessing data becomes much faster and this might (but not always will) result in abnormal acceleration.

Problem

What is an example of a parallel acceleration anomaly? ie. A task that when run over p processors, results in a speedup greater than p.

Original source