Performance, memory efficiency and programmability: the ambitious triptych of combining vertex-centricity with HPC
Capelli, Ludovic A. R.
The field of graph processing has grown significantly due to the flexibility and wide applicability of the graph data structure. In the meantime, so has interest from the community in developing new approaches to graph processing applications. In 2010, Google introduced the vertex-centric programming model through their framework Pregel. This consists of expressing computation from the perspective of a vertex, whilst inter-vertex communications are achieved via data exchanges along incoming and outgoing edges, using the message-passing abstraction provided. Pregel ’s high-level programming interface, designed around a set of simple functions, provides ease of programmability to the user. The aim is to enable the development of graph processing applications without requiring expertise in optimisation or parallel programming. Such challenges are instead abstracted from the user and offloaded to the underlying framework. However, fine-grained synchronisation, unpredictable memory access patterns and multiple sources of load imbalance make it difficult to implement the vertex centric model efficiently on high performance computing platforms without sacrificing programmability. This research focuses on combining vertex-centric and High-Performance Comput- ing (HPC), resulting in the development of a shared-memory framework, iPregel, which demonstrates that a performance and memory efficiency similar to that of non-vertex- centric approaches can be achieved while preserving the programmability benefits of vertex-centric. Non-volatile memory is then explored to extend single-node capabilities, during which multiple versions of iPregel are implemented to experiment with the various data movement strategies. Then, distributed memory parallelism is investigated to overcome the resource limitations of single node processing. A second framework named DiP, which ports applicable iPregel ’s optimisations to distributed memory, prioritises performance to high scalability. This research has resulted in a set of techniques and optimisations illustrated through a shared-memory framework iPregel and a distributed-memory framework DiP. The former closes a gap of several orders of magnitude in both performance and memory efficiency, even able to process a graph of 750 billion edges using non-volatile memory. The latter has proved that this competitiveness can also be scaled beyond a single node, enabling the processing of the largest graph generated in this research, comprising 1.6 trillion edges. Most importantly, both frameworks achieved these performance and capability gains whilst also preserving programmability, which is the cornerstone of the vertex-centric programming model. This research therefore demonstrates that by combining vertex-centricity and High-Performance Computing (HPC), it is possible to maintain performance, memory efficiency and programmability.