Objectives
Study how to develop compute- and data-intensive e-Science applications on large-scale hybrid distributed systems, taking performance and energy-consumption into account.
Background
Future high-performance and distributed systems (clusters, Grids, Clouds) will be far more diverse than today’s systems, as new types of processor accelerators are becoming mainstream. Examples include multi-core processors (e.g. AMD’s 48-core MagnyCours), Graphics Processing Units (GPUs), and Field Programmable Gate Arrays (FPGAs). Apart from the need for speed and scalability for many e-Science applications, there is increasing concern about power consumption.
Description
Develop a distributed programming environment by extending the successful Ibis system with support for accelerators, by integrating multiple equivalent compute kernels (‘equi-kernels’) each targeted at one specific accelerator. Implement applications from e-Science applications in COMMIT and elsewhere using the programming environment. Do controlled experiments on DAS-4 and large-scale experiments on other Grids and Cloud systems. Integrate the scheduling algorithms developed in this Project into the environment, allowing applications to make tradeoffs between performance and energy consumption.