Shared Scratch-Pad Memory Space Management

Abstract

Scratch-pad memories (SPMs) are important storage components in many embedded applications and used as an alternative or a complimentary storage to on-chip cache memories. One of the most critical issues in the context of SPMs is to select the data elements to place in them since the gap between SPM access latencies and off-chip memory access latencies keep increasing dramatically. Previous research considered this problem and attacked it using both static and dynamic schemes. Most of the prior efforts on data SPMs have mainly focused on single application scenarios, i.e., the SPM space available is assumed to be managed by a single application at any given time. While this assumption makes sense in certain domains, there also exist many cases where multiple applications need to share the same SPM space. This paper focuses on such a multi-application scenario and proposes a nonuniform SPM space partitioning and management across concurrently-executing applications. In our approach, the amount of data to be allocated to each application is decided based on the data reuse each application exhibits.