GCN-RA: A graph convolutional network-based resource allocator for reconfigurable systems

Abstract

Nowadays, hardware architectures with various reconfiguration capabilities provide significant computational speedup using parallelism and concurrency features. However, data transmission after allocating resources to the application is one of the critical challenges that produces communicational delays and time overheads, specifically in the execution of large-scale applications. This paper proposes a novel 2-level inter and intra-cluster resource allocation approach to reduce communication distances by defining regional resources. The requested application is partitioned through a customized Graph Convolutional Network (GCN) to achieve high-quality, independent parts with the lowest overhead and map them to adjacent or non-adjacent regions with an analytical distance metric. Using this approach, it is possible to efficiently obtain the general configuration of the final map solution by ignoring long hop-count distances and improving the convergence rate of the optimization algorithms. Many experiments on large-scale synthetic and real applications derived from CAD flow have been performed to evaluate the effectiveness of the proposed approach in comparison with the previous works. The results showed that in fixed optimization iterations, up to 15.97 % improvement in mapping solution quality has been achieved, and the resource utilization factor has reached 96.75 % value.