Scratch-Pad Memory Allocation without Compiler Support for Java Applications

ABSTRACT This paper presents the first scratch-pad memory allocation scheme that requires no compiler support for interpreted-language based applications. A scratch-pad memory (SPM) is a fast compilermanaged SRAM that replaces the hardware-managed cache. Its uses are motivated by its better real-time guarantees as compared to cache and by its significantly lower overheads in energy consumption, area and access time. Interpreted languages are languages such as Java that are interpreted by a run-time environment instead of being executed directly on hardware. All existing memory allocation schemes for SPM require compiler analysis to develop the allocation strategy. Specifically, existing allocation schemes for Java-based applications determine the allocations at compile-time. They then annotate the Java bytecodes with these allocation decisions for the Java Virtual Machine (JVM) to implement the actual allocation at runtime. These existing allocation schemes tie the resulting bytecode to specific SPM sizes, therefore preventing the applications from being portable to different SPM sizes. Further, existing methods do not work for unmodified third-party bytecodes produced by compilers other than their specialized compilers. In this paper, we propose the first ever SPM allocation scheme that is completely implemented inside the JVM. Our method requires no compiler support and works for unmodified bytecodes from any source. Moreover, unlike existing methods, it preserves the portability of bytecode to any SPM size. We investigate our method on the Sun Hotspot JVM on which we achieve a 27.8 % improvement on runtime and 21.8 % on energy saving versus not using the SPM – the only existing alternative for unmodified bytecodes.