Object-orientation as an appropriate paradigm for high-performance environments for scientific computing

Abstract

The scientific computing community is searching for high-performance environments to run their applications. These applications come from a variety of domains and have one aspect in common: they require many CPU-cycles. Examples of such applications areas are: computational fluid dynamics, financial modeling, weather forecasting, computational chemistry, pharmaceutical design, seismic data analysis, reservoir modeling, structural analysis and engineering design in the automotive and aerospace industries.Parallel computers with distributed memory are one type of environments which are being used in quite a few places for these applications. However, writing applications for these systems turns out to be a non trivial problem. One of the reasons is the limited support that is currently being offered on the available commercial or experimental systems. When talking to users of these systems (currently the users are also the programmers), it seems that another reason should not be underestimated. The users are very much concerned about execution efficiency; they are afraid of operating systems because they see them as too 'general', and feel they probably waist a large amount of the CPU cycles that could be used by the applications. For this reason, code for system services (e.g. message passing, routing, scheduling and load balancing) are often integrated in the application, which runs on a minimal software support environment. This usually leads to significant performance improvements in the short run, but is not feasible for the more complex applications.The challenge for the operating system community in this area is then to build execution environments adapted to the specific needs of these applications.