fat-tree topology

**Fat-tree topology** is the **network architecture with increasing uplink capacity toward the core to maintain high aggregate throughput** - it is commonly used in HPC and AI clusters where many nodes require simultaneous high-bandwidth communication. **What Is Fat-tree topology?** - **Definition**: Hierarchical switched fabric where higher tree levels are provisioned with wider bandwidth links. - **Design Goal**: Prevent core bottlenecks and support near non-blocking communication patterns. - **AI Cluster Fit**: Works well for collective-heavy workloads needing strong all-to-all communication behavior. - **Deployment Variables**: Oversubscription ratio, switch radix, cable plan, and expansion strategy. **Why Fat-tree topology Matters** - **High Throughput**: Sufficient core capacity preserves performance under concurrent multi-job traffic. - **Predictable Latency**: Balanced tree design reduces congestion hot spots and queueing spikes. - **Scalability**: Supports structured growth while retaining known performance properties. - **Collective Performance**: Strong bisection capacity benefits all-reduce and parameter exchange phases. - **Operational Visibility**: Hierarchical layout simplifies monitoring and fault-domain isolation. **How It Is Used in Practice** - **Capacity Planning**: Size spine and aggregation links for expected worst-case east-west traffic. - **Oversubscription Policy**: Set target oversubscription ratio based on workload sensitivity and budget. - **Validation**: Benchmark bisection and collective behavior after deployment and each expansion phase. Fat-tree topology is **a proven network pattern for communication-intensive AI infrastructure** - adequate uplink width at higher tiers is essential to avoid hidden scaling bottlenecks.