Understanding the Design-Space of Sparse/Dense Multiphase GNN dataflows on Spatial Accelerators

TL;DR

This paper characterizes the design space of dataflows for running Graph Neural Networks on spatial accelerators, focusing on the interplay of dense and sparse phases to optimize performance.

Contribution

It introduces a comprehensive taxonomy for mapping GNN phases on spatial accelerators, enabling better design and optimization of dataflows for GNN inference.

Findings

Different dataflow choices impact performance and efficiency.

Hardware parameters influence the effectiveness of dataflows.

Flexibility in dataflow supports pipelined execution and improves throughput.

Abstract

Graph Neural Networks (GNNs) have garnered a lot of recent interest because of their success in learning representations from graph-structured data across several critical applications in cloud and HPC. Owing to their unique compute and memory characteristics that come from an interplay between dense and sparse phases of computations, the emergence of reconfigurable dataflow (aka spatial) accelerators offers promise for acceleration by mapping optimized dataflows (i.e., computation order and parallelism) for both phases. The goal of this work is to characterize and understand the design-space of dataflow choices for running GNNs on spatial accelerators in order for mappers or design-space exploration tools to optimize the dataflow based on the workload. Specifically, we propose a taxonomy to describe all possible choices for mapping the dense and sparse phases of GNN inference, spatially and temporally over a spatial accelerator, capturing both the intra-phase dataflow and the inter-phase (pipelined) dataflow. Using this taxonomy, we do deep-dives into the cost and benefits of several dataflows and perform case studies on implications of hardware parameters for dataflows and value of flexibility to support pipelined execution.