Table of Contents

Interconnect

Interconnect is the network fabric connecting compute nodes (or sockets within a node) in a parallel system. It moves data between processors and is characterized by two independent metrics: latency (fixed time to get any message across) and bandwidth (sustained rate of large transfers).

Workloads are either latency-bound (many small messages) or bandwidth-bound (large bulk transfers), and interconnect optimization focuses on the dominant regime.

Example

Fat tree vs torus topology trade-offs.

Fat tree:    hierarchical switches, flat worst-case latency, higher cost
Torus/Mesh:  direct nearest-neighbor links, scales cheaply, hop count grows with size