Zhe Ma
PhD Student
IMEC (also K.U.Leuven/ESAT)
IMEC/DESICS
Kapeldreef 75, Leuven 3001, Belgium
My Home page
Exploration of interleaving tasks for power-efficient mapping of multi-media applications
The trend that ambient intelligence systems are emerging in the post-PC era has lead to the collision of two worlds: the classical large-scale software world and the high performance real-time embedded systems world. A typical unwanted scenario, which exists today, is that a software designer designs, in an object-oriented manner, large and reusable legacy code without considering the fact that his code will (partially) be executed on an energy and cost sensitive embedded hand-held device. The existing Task Concurrency Management (TCM) approach at IMEC will provide a concurrency optimization of the software for the target embedded platform that is performed at both design-time and run-time.
Today's embedded software becomes much larger, and contains large portions of static software components such as the multimedia application codecs. The existing TCM design-time scheduling cannot deal with very large systems efficiently. The PhD research work at design-time is to develope a scalable design-time scheduling process which can handle very large system specification (e.g. with hundreds of threads within a task) without significant loss of scheduling quality.
At run-time, due to the impact of memory access contentions, the external timing contraints and the lack of thread level parallelism by its nature, many tasks do not contain sufficient usable parallelism to fill all the processor simultaneously. Previous TCM runtime scheduling cannot handle this issue and hence this lower parallelism will result in a higher overall energy consumption to execute a given application task set, especially the static contribution. Therefore it is a key phase in the TCM workflow that the tasks without enough execution concurrency should be overlapped and interleaved to achieve higher execution concurrency. This will be reflected in the automatically generated middelware on top of RTOS to steer the overlapping of multiple tasks. The PhD research work mainly focuses on the dynamical interleaving algorithm and the integration of the interleaving phase into the existing TCM workflow.