論文ID: 21.20240460
Non-volatile memory (NVM) technologies (e.g., phase change memory (PCM), magnetic random access memory (MRAM), etc) are projected to be able to provide higher capacity with lower cost than DRAM in the near future. Accordingly, recent researchers are attempting to construct heterogeneous memory systems by combining NVM with DRAM to overcome the disadvantages of DRAM. However, the different characteristics of NVM and DRAM pose a major challenge, which carefully places the data in the appropriate memory without the requirement of disruptive change to applications as far as possible.
In this paper, we propose a compiler-assisted data placement technique for heterogeneous memory systems, including NVM and DRAM. Our scheme exploits the features of the compiler to enable data placement without the modification of applications or the OS. With the assistance of the compiler, we collect information (e.g., memory usage) for dynamic placement and allocation of data objects between NVM and DRAM based on the information. We implement our scheme in a Low Level Virtual Machine (LLVM) compiler infrastructure and evaluate its performance in real systems by injecting NVM latency using Quartz. In the experimental results, we demonstrate the effectiveness of our scheme for different memory configurations.
