Virtualization of Computing Resources in RCS for Multi-task Stream Applications

Abstract

The possibility for distribution of FPGA resources in the temporal domain for multi-modal & multi-task workloads conceptually allows virtualization of logic, communication and input/output resources similar to memory virtualization in advanced conventional computers (e.g. superscalar). This, in turn, can dramatically increase the cost-effectiveness of FPGA based Reconfigurable Computing Systems (RCS). In the presented “proof-of-concept” research the following topics have been investigated, developed and tested: i) architecture of a platform to support the dynamic allocation of Application Specific Virtual Processors (ASVP), ii) mechanisms for run-time on-chip assembly of ASVP from Virtual Hardware Components (VHC) and iii) mechanisms for run-time on-chip components (VHC) relocation in predetermined regions of the FPGA device. The above mechanisms have been implemented and tested on a specially developed platform: the Multi-task Adaptive Reconfigurable System (MARS) Platform. The actual application of MARS was prototyping a high-performance multi-mode stereo-vision system (200 fps) for the next generation of space-borne computing platforms.