Grid-based Image Registration

We introduce and discuss preliminary experience with an application that has vast potential to exploit the Grid for social benefit and offers interesting resource assessment and allocation challenges, having real-time aspects: image registration. Image registration is generally formulated as an optimization problem that satisfies constraints, such as coordinate displacements that are affine or volumepreserving or that obey the laws of elasticity. Three-dimensional registration of high-resolution images is computationally complex and justifies parallel implementation. In turn, ensembles of registration tasks exploit concurrency in the simpler sense of job farming. Registration is an elementary example of a much larger class of large-scale mesh-based computations that are in principle amenable to execution on the Grid, but are sensitive to workload-to-capability balance at synchronization points. While better resource assessment and allocation tools lift all such applications, reducing sensitivity to synchronization within an individual application is a complementary and equally important objective. We therefore examine the potential for weakening the synchronization sensitivity of general mesh-based bulk synchronous computations through less restrictive programming paradigms.