Heat-aware Graph Data Placement Strategy for NVM
Rongding Sheng · 2023
As a data-intensive application, graph processing has poor locality and many irregular memory accesses, which lead to poor processing performance on traditional memory architectures. Non-volatile memory (NVM) has the potential to replace DRAM as a large-capacity memory for computing devices due to its high storage density, low power consumption, and byte addressability. Based on NVM, this paper proposes a cold and hot data placement strategy (CHDPS) combining hardware and software. Specifically, we first explored the memory access modes of hot and cold data, and found that hot data has performance defects such as high access frequency and large memory access delay; then combined with the two-level buffer inside NVM, the hot and cold data are relocated, the hot data is stored in the more efficient RMW buffer, the cold data is placed in the inefficient AIT Buffer, and the Cacheline of the AIT Buffer is partitioned at the same time; finally this strategy was implemented in the Gem5+VANS simulation environment, and the performance was improved by an average of 33% and 5.1% compared with the NVM system and the hybrid memory system, respectively. Compared with the traditional DRAM system, the performance was close to 91%.