Topology-Aware Load-Balance Schemes for Heterogeneous Graph Processing

Inspired by the insights presented in Chapters 2, 3, and 4, in this chapter the authors present the KCMAX (K-Core MAX) and the KCML (K-Core Multi-Level) frameworks: novel k-core-based graph partitioning approaches that produce unbalanced partitions of complex networks that are suitable for heterogeneous parallel processing. Then they use KCMAX and KCML to explore the configuration space for accelerating BFSs on large complex networks in the context of TOTEM, a BSP heterogeneous GPU + CPU HPC platform. They study the feasibility of the heterogeneous computing approach by systematically studying different graph partitioning strategies, including the KCMAX and KCML algorithms, while processing synthetic and real-world complex networks.

Core Kernels for Complex Network Analysis

The emergence of Network Science has motivated a renewed interest in classical graph problems for the analysis of the topology of complex networks. For example, important centrality metrics, such as the betweenness, the stress, the eccentricity, and the closeness centralities, are all based on BFS. On the other hand, the k-core decomposition of graphs defines a hierarchy of internal cores and decomposes large networks layer by layer. The k-core decomposition has been successfully applied in a variety of domains, including large graph visualization and fingerprinting, analysis of large software systems, and fraud detection. In this chapter, the authors review known efficient algorithms for traversing and decomposing large complex networks and provide insights on how the decomposition of graphs in k-cores can be useful for developing novel topology-aware algorithms.

A Case Study on Citation Network Analysis

In this chapter, the authors present a case study of Network Analysis in the field of bibliometrics, focused on the identification of central academic articles based on complex network metrics that can be implemented with algorithms covered throughout this book. The authors analyze a scientific citation network and systematically obtain the most central papers considering different perspectives of the selected document collection. Later, they discuss the potential benefits that the parallel kernels and the topology-aware partitioning algorithms can offer in the context of the presented study case. Finally, the authors summarize this book's main contributions and offer some concluding remarks.

Partitioning of Complex Networks for Heterogeneous Computing

The most fundamental problem in BSP parallel graph computing is to decide how to partition and then distribute the graph among the available processors. In this regard, partitioning techniques for BSP heterogeneous computing should produce computing loads with different sizes (unbalanced partitions) in order to exploit processors with different computing capabilities. In this chapter, three major graph partitioning paradigms that are relevant to parallel graph processing are reviewed: balanced graph partitioning, unbalanced graph partitioning, and community detection. Then, the authors discuss how any of these paradigms fits the needs of graph heterogeneous computing where the suitability of partitions to hardware architectures plays a vital role. Finally, the authors discuss how the decomposition of networks in layers through the k-core decomposition provides the means for developing methods to produce unbalanced graph partitions that match multi-core and GPU processing capabilities.

Accelerated Network Traversal Using Multi/Many-Cores

New BSP platforms like TOTEM are following a similar approach to Big Data industry-proven frameworks but applied to exploit the potential of HPC heterogeneous nodes that combine multi-cores and GPUs. The performance of platforms under this new paradigm depends on minimizing the computation time of partitions by increasing the suitability of partitions to processors. However, there is a lack of studies on the suitability of parallel architectures for processing different families of graphs, including small-world and scale-free networks. In this chapter, the authors show how to characterize the performance of multi/many-cores when traversing synthetic networks of varying topology in order to reveal the suitability of multi-cores and GPUs for processing different families of graphs.

The Need for HPC Computing in Network Science

The size of complex networks introduces large amounts of traversal times that can be tackled by exploiting pervasive multi-core and many-core parallel hardware architectures. However, there is a list of factors that make the design of efficient parallel traversal algorithms for graphs difficult: unstructured problems, data-driven computation, irregular memory access, poor locality, and low computing load. In this chapter, the authors introduce the synergy between Network Science and High Performance Computing and motivate the combined use of multi/many-core heterogeneous computing and Network Science techniques to tackle the above-mentioned challenges and to efficiently traverse the structure of massive real-world graphs.

Trends and Challenges in Large-Scale HPC Network Analysis

Many algorithms in graph analytics can be sped up by using the power of low-cost but massively parallel architectures, such as GPUs. On the other hand, the storage and analysis capabilities needed for large-scale graph analytics have motivated the development of a new wave of HPC technologies, including MapReduce-like BSP distributed analytics, No-SQL data storage and querying, and homogeneous and hybrid multi-core/GPU graph supercomputing. In this chapter, the authors review these trends and current challenges for HPC large-scale graph analysis.