Towards dataflow based graph processing

While existing researches greatly improve the performance of memory subsystem [4], they are still subject to the underlying modern processor. We divide the process of graph applications into three pieces of slots, which indicates hardware resources needed to process micro-ops (uops), and present the result in Figure 1(a). Although previous work has made great progress to significantly improve the performance of graph applications by optimizing 41% stalls resulting from the memory subsystem [5,6], a vast body of inefficiencies (35% slots wasted) inside the underlying processor is still unknown and seldom studied in existing work.