Task Scheduling for Adaptive Reconfigurable VLIW Multicore Processors

Abstract

Embedded Reconfigurable Architectures (ERA) is a project with the objective to design a platform that combines reconfigurable computing and network elements which can adapt on-the-fly their composition, organization and even instruction-set architectures in an effort to provide the best possible trade-offs in performance and power for the given application(s). Although some of this adaptiveness is controlled by software (mainly the operating system), the goal of the ERA project is that great deal of this control actually takes place automatically at hardware level, by the Hardware scheduler. This thesis deals specifically with the problem of hardware task scheduling. We studied a variety of possible implementations for the task scheduling in the ERA platform. After getting an inside look of the Processing component of the ERA platform and understanding the particularities of it, as well as of its main building block, the ?-VEX core, we tried to find scheduling algorithms available in the bibliography to implement as the Hardware scheduler of ERA. This literature research did not yield any results, both because of the complexity of the problem, as well as the pioneer characteristics of ERA that we would like to take advantage of. We designed some simple scheduling algorithms, especially tailored for the ERA platform and tested them. The most important of them were Basic, which simply stalls the tasks until there are enough resources for them, Versioning, which brings a different binary from the memory which is compiled to run on a smaller core and Generic Binary which uses a binary that is especially compiled to run on any core, so downgrading a task does not lead to the communication cost that Versioning suffers and upgrading a task becomes possible. We present the most important of the experiments that took place within this thesis and show that GB++ (a version of GB that supports forced priorities, interrupts and upgrading by default) is the most promising algorithm that can take advantage of all the characteristics and the abilities of ERA, without being the fastest one, which is Versioning. Finally, we researched and defined the minimum requirements of GB++ in order to become apart from the rest also the fastest algorithm for ERA.