Modeling large-scale platforms for the analysis and the simulation of scheduling strategies

Summary form only given. An important trend in scientific computing is the establishment of computing platforms that span multiple institutions to support applications at unprecedented scales and levels of performance. A key issue for achieving high performance is the scheduling of application components onto available resources, which has been an active area of research for several decades. However, most of the platform models traditionally used in scheduling research, and in particular the network models, break down for platforms spanning multiple (wide-area) networks. In this paper we examine modeling issues for large-scale platforms. More specifically, we discuss network latency, bandwidth sharing, and network topology. Our discussion is from the perspective of scheduling research and the main challenge we address is to develop models that are sophisticated enough to be realistic, but simple enough that they are amenable to analysis. Finally, while the models we propose can be used to study scheduling problems directly, they also form a good basis for realistic simulation, which is often the method of choice for comparing scheduling strategies.