A Semi-automatic Scratchpad Memory Management Framework for CMP

Abstract

Previous research has demonstrated that scratchpad memory(SPM) consumes far less power and on-chip area than the traditional cache. As a software managed memory, SPM has been widely adopted in today's mainstream embedded processors. Traditional SPM allocation strategies depend on either the compiler or the programmer to manage the small memory. The former methods predict the frequently referenced data items before real running by static analysis or profiling, whereas the latter methods require the programmer to manually allocate the SPM space. As for the dynamic heap data allocation, there is no mature allocation scheme for multicore processors with a shared software-managed on-chip memory. This paper presents a novel SPM management framework, for chip multiprocessors (CMP) featuring partitioned global address space (PGAS) SPM memory architecture. The most frequently referenced heap data are maintained in the SPM. This framework mitigates the SPM allocation problem by leveraging the programmer's hints to determine the data items allocated to the SPM. The complex and error-prone allocation procedure is completely handled by an SPM management library (SPMMLIB) without programmer's conscious. The performance is evaluated in a homogenous UltraSPARC multiprocessor using PARSEC and SPLASH2 benchmarks. Experimental results indicate that, on average, the energy consumption is reduced by 22.4% compared with the cache memory architecture.