Mitigating Scalability Walls of RDMA-based Container Networks
summary
ScalaCN infers RNIC internals via combinatorial causal testing and links specific config dimensions to failures, finding that increasing flow-group query cost (not entry count) collapses throughput (examples show RNIC bandwidth drops up to 87% when containers scale 0.4M→0.8M and some NICs go unresponsive past ~7,000 VxLAN contexts). Applying a multi-level flow-group cache and reordering offloads raises end-to-end bandwidth 1.4×, cuts latency 31%, and resolves 82% of inferred causes.
AI-written summary · about
chapters (12)
- 0:00RDMA container networks intro
- 1:41Scalability challenges in RCNs
- 3:08Root causes in ANICs
- 6:16Black-box ANIC limits
- 7:41Scala for ANIC testing
- 9:25Causal testing methodology
- 10:31Flow group query impact
- 11:58Kernel driver failures
- 12:41Performance prediction models
- 13:23Scara flow offloading
- 14:29Scara evaluation impact
- 14:50Summary and future
related talks
34mTroubleshooting Low Latency Application on CNF Deployment
26mFreeFlow: Software-based Virtual RDMA Networking for Containerized Clouds
25mRDMA @Scale
30mThe Gashlycrumb Tinies of AI Networking You Must Know (or Languish!)
20mRDMA at Cloud Scale: The OCI Experience
34mComprehensive Performance Benchmark on Various Well-Known CNI Plugins
13mContracting Wide-area Network Topologies to Solve Flow Problems Quickly
31mLive Panel: Scale-Out Networking for AI Supercomputers