CoSPARSE: A Software and Hardware Reconfigurable SpMV Framework for Graph Analytics

Abstract

Sparse matrix-vector multiplication (SpMV) is a critical building block for iterative graph analytics algorithms. Typically, such algorithms have a varying active vertex set across iterations. This variablity has been used to improve performance by either dynamically switching algorithms between iterations (software) or designing custom accelerators (hardware) for graph analytics algorithms. In this work, we propose a novel framework, CoSPARSE, that employs hardware and software reconfiguration as a synergistic solution to accelerate SpMV-based graph analytics algorithms. Building on previously proposed general-purpose reconfigurable hardware, we implement CoSPARSE as a software layer, abstracting the hardware as a specialized SpMV accelerator. CoSPARSE dynamically selects software and hardware configurations for each iteration and achieves a maximum speedup of 2.0× compared to the naïve implementation with no reconfiguration. Across a suite of graph algorithms, CoSPARSE outperforms a state-of-the-art shared memory framework, Ligra, on a Xeon CPU with up to 3.51× better performance and 877× better energy efficiency.

MeNDA

“…Since Beamer et al [5] first proposed a hybrid approach for Breadth First Search, many recent graph analytics frameworks have built upon this work and adopted dynamic reconfiguration between a sparse and a dense representation of the dataflow based on the active vertex set [9,13,17,21,33,37,43,48,50,56,59]. The dynamic reconfiguration greatly improves performance but requires the original graph 𝐴 for one representation and its transpose 𝐴 𝑇 for the other representation during execution.…”

“…2(a)) is that recent breakthroughs in algorithms and architectures have significantly improved the performance of graph processing. Consequently, runtime transposition using a state-of-the-art implementation [49] can introduce a 126% performance overhead to a recently proposed graph framework [17]. Therefore, graph frameworks usually store more than one copy of the input graph in different formats to avoid the performance overhead of transposing the graph on-the-fly.…”

“…Prior work proposed two algorithms for parallel sparse matrix transposition -a count sort based algorithm (scanTrans) and a merge sort based algorithm (mergeTrans) [49]. In this work, we adopted the merge sort algorithm not only because merge sort presents higher spatial locality but also because merge sort is widely used in sparse linear algebra [17,38,42]. Inspired by prior near-DRAM accelerators [4,24], the near-memory processing units (PUs) are embedded in the buffer chips of DIMMs to minimize the modifications to commodity DRAM devices.…”

“…As a fundamental primitive in many important application domains such as graph analytics, machine learning, and scientific computation [3,8,17,20,22,38,41,45,47,55,58], Sparse Basic Linear Algebra Subprograms (SpBLAS) are notoriously memory intensive due to the irregular memory access pattern. Recently, there has been a surge in customizing hardware accelerators near memory to tackle sparse BLAS applications such as sparse gathering [2,24,30], sparse matrix vector multiplication (SpMV) [2,42,52], and graph analytics [1,12,36,57,60].…”

“…Merge sort is widely employed in sparse linear algebra applications, making MeNDA an efficient solution for many sparse dataflows. Finally, to showcase its applicability to other sparse dataflows, we illustrate how MeNDA can be adapted to perform SpMV, which is a fundamental kernel for machine learning and graph analytics [2,17,38,42].…”