Deadlock avoidance in non-minimal routing for high-radix networks

Abstract

Avoiding routing deadlocks is an important component of interconnection networks that prevents the network from progressing due to cyclical dependency within the networks. For large-scale systems using high-radix topologies that leverage non-minimal adaptive routing, virtual channels(VCs) are generally used to avoid routing deadlock. However, VCs can be costly for large-scale networks due to the deep buffers. In this work, we propose both routing algorithms and flow controls -- Restricted Intermediate-node Non-minimal Routing (RINR) algorithm and opportunistic flow control (OFC), which both exploit the low-diameter characteristics of high-radix networks while still maximizing path diversity within the topology. We will refer these algorithms as BoomGATE. We analyze the occurrence of routing deadlock in fully-connected topologies due to the non-minimal routes, and propose to restrict the non-minimal routes to guarantee deadlock freedom without any additional VCs. We also propose an algorithm to ensure path diversity that is able to load-balanced across all nodes in the network. However, since path diversity is restricted with the RINR algorithm, we propose opportunistic flow control (OFC) to complement the path diversity limitation of RINR. OFC is the flow control method to allow the "illegal routes" if and only if sufficient buffer can be guaranteed to ensure cyclical dependency does not occur. We propose both a static and dynamic OFC implementation. We evaluate the performance of BoomGATE and demonstrate that there is minimal performance loss compared to global adaptive routing, while achieving 50% reduction of the amount of buffers.