Overcome Interconnect Bottlenecks in HPC Clusters
summary
Claims and measured wins focus decision making not topic labels: GPU clusters suffer 30–50% execution overhead from interconnect latency, RDMA promises 55% end-to-end latency reduction and 97% protocol efficiency for mid‑sized messages, Frontier reports a ~24.8% throughput uplift with Slingshot and adaptive routing, and Nvidia shows NVLink 50 GB/s at 1.8 µs with NCCL topology‑aware collectives. The talk lists concrete topologies, benchmark numbers, and software tactics and reports whether those changes actually reduce stalls at scale.
AI-written summary · about
chapters (10)
- 0:00Introduction and Welcome
- 1:25Traditional Interconnect Limitations
- 2:23Innovative Network Topologies
- 3:30RDMA and Efficiency Gains
- 4:25Case Studies: Frontier and Nvidia Seline
- 6:08Software Optimization Strategies
- 7:25Future Interconnect Technologies
- 8:20Performance Metrics and Benchmarks
- 9:09Key Takeaways and Next Steps
- 9:55Conclusion and Contact Information
related talks
20mBenchmarking HPC workloads on OpenShift
34mSupercharge Kubeflow Performance on GPU Clusters
25mHigh Performance Networking for Distributed DL Training in Production K8s
09m⚡ Lightning Talk: Advanced GPU-Orchestrated Workflows and HPC Integrations on K8s for Distributed AI/ML at Scale
30mGPU Sharing at CERN: Cutting the Cake Without Losing a Slice!
22mBeyond Single-Cluster Limits: Scaling GPU Workloads Across Kubernetes With…
34mRDMA Enabled Kubernetes for High Performance Computing
21mExperience Operating Large GPU Clusters at Organizational Scale