Modern computer applications, especially embedded applications, show an increased demand for hardware acceleration to cope with changing required computation power while adhering to given constraints such as energy consumption and battery life.

With the convergence of formerly separated fields of applications such as e.g. telecommunication, networking, and multimedia such devices become increasingly complex and therefore power-hungry. Using standard design techniques, modern embedded devices are typically designed around an embedded general purpose processor enhanced by dedicated accelerators. With an ever increasing number of formats, protocols, and algorithms to support, this leads to either over-designed architectures or hardware-support for a limited subset of algorithms used within the individual applications. These applications are then tuned to run most efficienctly on the resulting architecture.

Reconfigurable architectures enable fully dynamic systems: instead of fixed hardware accelerators field programmable logic can be embedded which will be configured on-demand with respect to application demands and architecture constraints.

In this proposal, we want to focus on reconfigurable architectures with emphasis on self-optimization. This is achieved as follows: a dedicated monitoring infrastructure is used to determine existing or potential bottlenecks. Based on this information decisions regarding system optimization are made. Optimization may follow the classic approach, i.e. guided off-line optimization, or invocation of pre-fabricated accelerator modules. Monitoring data may also be fed back into the design process to
better guide application compilation and accelerator synthesis.

Based on our work on semi-automated integrated hardware-software codesign, we also want to explore the possibilities of on-demand accelerator synthesis and on-line application modification.

Key to this approach is research in and development of novel monitoring techniques which enable proactive system monitoring and autonomous data correlation in addition to traditional reactive, role-based monitoring. We want to explore how this information can be used to direct off-line and on-line system reconfiguration. For this we need to identify appropriate metrics, i.e. which monitoring data is required to achieve required system evaluation to finally guide system reconfiguration.

Cluster meetings between our two chairs (at least one per quarter), 14-day exchange of Ph.D. students.