A High-Performance Design, Implementation, Deployment, and Evaluation of The Slim Fly...
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
Builds and validates the first real-world Slim Fly cluster: a 200-server, 50-router installation with practical cabling patterns and an InfiniBand routing architecture that uses almost-minimal, disjoint-path selection across service levels. Uses algebraic switch IDs and automated checks for miswired ports to speed deployment. Benchmarks include deep neural net training, graph analytics, and linear algebra kernels; Slim Fly shows better scalability and comparable-or-better throughput and lower cost versus a non-blocking Fat Tree baseline. Open-source routing stack is fully portable.
AI-written summary · about
chapters (8)
- 0:00Slim Fly introduction
- 1:01Why Slim Fly matters
- 2:05What is Slim Fly
- 5:16Deploying Slim Fly
- 10:56Routing challenges
- 13:45Solving routing paths
- 16:12Slim Fly vs Fat Tree
- 18:39Key findings
related talks
27mSlim: OS Kernel Support for a Low-Overhead Container Overlay Network
26mF10: A Fault-Tolerant Engineered Network
19mStarfish: A Topology-Routing Co-Design for Small-Scale Data Centers
23mBalancing on the Edge: Transport Affinity without Network State
32mSnabb Switch simple and fast packet networking
17mCo-Designing Traffic Control with NVMe-oF for Disaggregated Storage: A Comparative Study
24mNetBlocks: Staging Layouts for High-Performance Custom Host Network Stacks
14mHigh-throughput and Flexible Host Networking for Accelerated Computing