In this talk, we will look at the performance side of large scale data processing. We will look at distributed memory systems and see how the decisions taken by the middleware can have significant impact on the performance of the application. We will consider three different middlewares -- DataCutter (filter-stream), KAAPI (workstealing) and Map-Reduce MPI (Map Reduce) on three different applications -- blackscholes option pricing, medical image analysis and, synthetic aperture radar. Using these examples we will demonstrate that one needs to care about each part of a computation to reach the highest performance.
