Managing Reconfigurable Resources in Heterogeneous Cores Using Portable Pre-Synthesized Templates

Abstract

In this paper we consider multi-FPGAs, reconfiguration and system description portability as the processes of specifying and modeling a complete system before it is partitioned and committed to a style/flow of implementation. In the case of a high performance computing cluster employing FPGAs, the reconfigurable elements need to be dynamically re-allocated and reconfigured based on the prevailing workload at a given instance. We particularly target fast configuration and task migration in high performance computing systems, such as server farms. This paper presents a design flow, based in resource sharing across configurations on the reconfigurable devices, which can manage the resource allocation and reconfiguration quickly when resources need to be migrated between different applications. The resources shared across configuration are defined using an isomorphic-driven partitioning approach. This technique detects recurrent structures and produces a partitioned specification in which the identified clusters are instances of repeating templates in the original graph used to describe the input application. Two algorithms for regularity extraction were implemented. The performances of the algorithms were compared with regard to the size and number of regular structures identified by the two approaches, as well as the lime taken to perform their task.