High Performance Large Graph Analytics by Enhancing Locality

Graphs are widely used in a variety of domains for representing entities and their relationship to each other. Graph analytics helps to understand, detect, extract and visualize insightful relationships between different entities. Graph analytics has a wide range of applications in various domains including computational biology, commerce, intelligence, health care and transportation. The breadth of problems that require large graph analytics is growing rapidly resulting in a need for fast and efficient graph processing. One of the major challenges in graph processing is poor locality of reference. Locality of reference refers to the phenomenon of frequently accessing the same memory location or adjacent memory locations. Applications with poor data locality reduce the effectiveness of the cache memory. They result in large number of cache misses, requiring access to high latency main memory. Therefore, it is essential to have good locality for good performance. Most graph processing applications have highly random memory access patterns. Coupled with the current large sizes of the graphs, they result in poor cache utilization. Additionally, the computation to data access ratio in many graph processing applications is very low, making it difficult to cover the memory latency using computation. It is also challenging to efficiently parallelize most graph applications. Many graphs in real world have unbalanced degree distribution. It is difficult to achieve a balanced workload for such graphs. The parallelism in graph applications is generally fine-grained in nature. This calls for efficient synchronization and communication between the processing units. Techniques for enhancing locality have been well studied in the context of regular applications like linear algebra. Those techniques are in most cases not applicable to the graph problems. In this dissertation, we propose two techniques for enhancing locality in graph algorithms: access transformation and task-set reduction. Access transformation can be applied to algorithms to improve the spatial locality by changing the random access pattern to sequential access. It is applicable to iterative algorithms that process random vertices/edges in each iteration. The task-set reduction technique can be applied to enhance the temporal locality. It is applicable to algorithms which repeatedly access the same data to perform certain task. Using the two techniques, we propose novel algorithms for three graph problems: k-core decomposition, maximal clique enumeration and triangle listing. We have implemented the algorithms. The results show that these algorithms provide significant improvement in performance and also scale well