Load-balancing and object-orientation: a case study from mesoscale flow modelling

Parallel processing is a natural choice to achieve higher performance for many applications from the area of scientific computing. Among them, the models of atmospheric circulation attract public attention, since they can be used for the simulation of the impact of human activities on our environment. The paper promotes the use of object-oriented design techniques to implement large data parallel applications. Object-parallelism is a very natural means to express concurrency inherent in an application. Above this we feel, that in an object-oriented environment it is much easier to provide fairly general solutions for problems common to parallel processing. One of the challenging problems of parallel processing is load balancing. For a wider acceptance of parallel computers solutions must be offered by the parallel operating systems. Of course this cannot be done irrespective of the applications, when efficiency has to be taken into account, Here again an interface on object level is the appropriate choice to combine generality with efficiency. We present a case study for the object-oriented redesign of a simulation program. Based on an analysis of this application we show that its inherent load imbalances can be redressed by integration of a simple load balancing strategy. A performance improvement by 25% has been achieved