A fast profiler for compilation of multi-threaded applications on a hybrid memory system

Abstract

This paper proposes a fast memory access profiling tool called HMTool for multi-threaded applications. The proposed profiling tool extracts memory access patterns rapidly to optimize a multi-threaded application by allocating variables properly on a hybrid memory system with DRAM and non-volatile memory. To enhance the profiling performance, this paper proposes a static analysis and a caching technique for profiling. The proposed tool is implemented based on LLVM compiler framework. Experiment results show that the proposed profiling techniques improve the profiling performance by 5.1 times compared with an existing memory profiling tool.