Online Scratchpad Memory Management for Multi-core Systems

Abstract

Scratchpad Memory (SPM) a software-controlled on-chip memory has been increasingly widely used in embedded systems because it has less access energy and higher area density when compared to an ordinary cache In order to reduce energy consumption the architecture of multicore system has been proposed which replaces cache by SPM In this architecture tasks can access data on both local and remote SPMs Latency of accessing data on remote SPM is longer than local Therefore SPM allocation will affect execution time of tasks In addition SPM allocation method is an important issue in reducing energy consumption Existing methods for allocating SPM space require offline program profiling which could result in inefficient SPM utilization due to a different input to the program during runtime or unawareness of the SPM information in the runtime environment This paper proposes SAMOS a novel online method without offline profiling for allocating SPM space on multicore system SAMOS performs SPM partition according to the dynamic access behavior of the running tasks But the partition may bring about a longer execution time by accessing data on remote SPM Therefore to minimize execution time of each task SAMOS places more frequently used data in local and rarely frequently used data in remote SAMOS is implemented based on the cooperation of hardware and software The evaluation results show that SAMOS can reduce the energy delay product (EDP) by up to 57% (29% on average) compared to a contention aware SPM allocation policy (CASA) Moreover the area overhead of the hardware support is insignificant (about 1 8%)

Date of Award	2015 Mar 12
Original language	English
Supervisor	Da-Wei Chang (Supervisor)