Trace-Based Method For Big Data Memory Characteristics Research

Big data has exacerbated the so-called "memory wall" problem. To study the memory characteristics of big data applications has become an important issue in the high end computing community. In this paper, we propose a trace-based method based on the trace files generated by simulators, which captures memory access information in different memory hierarchies and aggregates information to get memory performance statistics. Simulations were conducted to research the impact of cache size and hardware prefetch on big data applications, and our trace-based method was used to obtain the desired memory performance metrics. Experimental results show that big data benchmarks are less sensitive to cache size than traditional benchmarks, and hardware prefetching is effective in improving L2 cache hit rate. In terms of memory access address range, big data benchmarks have wider address range and the address distribution is more irregular than traditional benchmarks.