SC21 Replay: Advanced Congestion Control for Addressing Network Bottlenecks
The interconnect is one of the most critical components in large-scale computing systems, and its impact on the performance of parallel applications increases with system and workload size. It is crucial for facilitating high performance IO that can keep up with the demands of HPC applications: latency, bandwidth, topology routers, protocols, and must be designed and implemented efficiently to realize the potential of HPC clusters. However, previous generation switch-intensive networks are inherently congestive and directly contribute to poor workload performance.
Advanced Congestion Control for Addressing Network Bottlenecks, originally presented by Rockport CTO Matt Williams at SC21, focuses on a next-generation interconnect with performance enhancements made possible through an innovative distributed switchless architecture that delivers advances in path diversity, deadlock-free routing and intelligent applications of ultra-high priority QoS. Matt’s presentation discusses advances in congestion protection, resiliency, and operational simplicity, and shows best practices for benchmarking to predict performance in production environments. Also included in this 30-minute video is a demonstration of the Autonomous Network Manager, Rockport’s solution for network management transparency and simplicity.