Accelerated Hybrid Approach for Spectral Problems Arising in Graph Analytics

Abstract

Data sets such as graphs are growing so rapidly that performing meaningful data analytics in reasonable time is beyond the ability of common software and hardware for many applications. In this context, performance and efficiency are primary concerns. The spectral analysis of real networks reflects such problematic. In this paper we present a solution based on Krylov methods which combines accelerators to increase the throughput of graphs traversals and latency oriented architectures to solve small problems. We focus on an hybrid acceleration of the implicitly restarted Arnoldi method which targets large non-symmetric problems with irregular sparsity pattern. The result of this cooperation is an efficient solver to compute eigenpairs of real networks. Moreover, this approach can be applied to other methods based on coarsening.