Understanding Lifecycle Management Complexity of Datacenter Topologies
summary
Awarded Best Paper! Paper defines concrete deployment and expansion metrics—number of switches, patch panels, bundle taps and expansion-step count plus rewires-per-patch-panel-rack—and measures Clos vs. expander (jellyfish). Findings: no topology dominates; Clos wins on bundling and patch-panel count, jellyfish uses fewer switches and fewer expansion steps. From those principles a new family, Fat Clique, is generated by a constraint search (32-port radix baseline) that maximizes in-rack links and regularity; it matches expander-like expansion (about 3–4 steps) while keeping Clos-like wiring simplicity. Code release planned.
AI-written summary · about
chapters (23)
- 0:00Intro
- 0:31Datacenter topology designs
- 1:56Lifecycle management of datacenter topologies
- 2:09Management complexity is important
- 3:37Lifecycle management overview
- 4:16Packaging
- 4:47Wiring complexity
- 5:05Cable bundling
- 6:33Role of patch panel in bundling
- 6:59Deployment complexity metrics
- 7:33A single expansion step complexity
- 8:36Topology comparison case study
- 9:10Importance of regularity
- 9:32Importance of maximizing intra-rack links
- 9:58Importance of fat edge
- 11:21Summary of case study
- 12:31Challenges
- 13:23Constraint-based search
- 14:33Evaluation Methodology
- 15:00FatClique has low deployment complexity
- 15:36FatClique has low expansion complexity
- 16:30Conclusions and Future work
- 19:26Spectral gap
related talks
17mFlattened Clos: Designing High-performance Deadlock-free Expander Data Center Networks...
21mLarry: Practical Network Reconfigurability in the Data Center
19mStarfish: A Topology-Routing Co-Design for Small-Scale Data Centers
10m10x Network Design Engineer
14mRDC: Energy-Efficient Data Center Network Congestion Relief with Topological
10m10x Network Design Engineer
14mEfficient Direct-Connect Topologies for Collective Communications
20mPhysical Network Design at Scale