Configuration sharing based reconfiguration overhead management for partially reconfigurable architectures

Abstract

Reconfigurable architectures are becoming one of the most promising solutions in computing domains. Reconfigurable architectures` reconfigurability and hard-wired logic pursuit two important goals of flexibility and performance. However, applications of reconfigurable architectures are sometimes limited due to a reconfiguration overhead, for which many hardware supports and software translation techniques are proposed. Among those techniques, an architectural support called a partial reconfigurability that reconfigures only a required portion of a device as applications change is the hot issue to reduce the run-time reconfiguration overhead in recent commercial and academic reconfigurable architectures. Many exploits the partial reconfigurability with a time-multiplexing scheme that implements multiple temporal partitions at one time and selects one of them. In the time-multiplexing, the common resources are reused again and again in multiple temporal partitions. To maximize the amount of common resources, many resource sharing techniques have been proposed. However, the time-multiplexing requires relatively large area to maintain all temporal partitions and to control multiple datapaths with multiplexers. This dissertation proposes another way to employ the partial reconfigurability to reduce the reconfiguration overhead. The key idea is to share configurations between consecutive temporal partitions and to reduce the size of configuration bitstreams. Compared to the time-multiplexing, the proposed scheme repeatedly loads configurations, and only keeps one configuration at a time, which results smaller area. The introduced configuration sharing is divided into two sub-problems of a resource sharing and a placement and routing. For the first sub-problem, this dissertation proposes a novel graph model for interconnection and operation sharing, and a resource sharing algorithm that is based on a traditional graph isomorphism. To effectively reuse resources ...