Recall: Reordered Cache Aware Locality Based Graph Processing

Sparse graph processing generates highly irregular Memory Access Patterns (MAP) which lack locality and result in poor cache performance. In this paper, we propose a novel graph ordering algorithm that addresses this problem. We observe that existing reordering algorithms primarily try to improve cache line utilization by enhancing spatial locality. They are oblivious to cache data reuse which reflects the temporal locality that MAP can possess. Our premise is that peak efficiency can be achieved by a graph order for which the resulting MAP exhibit both spatial and temporal locality. Therefore, we first introduce a new metric Profit, that quantifies cache data reuse leading to a heuristic pH that enhances temporal locality in the MAP of graph algorithms. Then we define a notion of dynamically matching MAP with cache contents in a way that jointly maximizes both cache data reuse and cache line utilization. To perform this joint optimization, we develop a Block Reordering algorithm which utilizes pH to rearrange blocks of consecutive nodes with high spatial locality. We evaluate our algorithm using 8 real world datasets and 4 representative graph algorithms. Experimental results show that graphs obtained by Block Reordering can achieve upto 2.3x speedup over the original graph order and consistently outperform the existing state of the art reordering technique by 20% to 25% reduction in cache misses.