Understanding RDMA Microarchitecture Resources for Performance Isolation
summary
Presents a model tying RDMA verbs to RNIC internal resources using hardware counters, then builds the Husky test suite that creates verb pairs to target specific resources. Finds memory registration thrashes NIC cache so a zero-bandwidth attacker cuts a victim from ~100 Gbps to ~10 Gbps, and receive-not-ready exhausts RX pipeline so a simple attacker drops a 97 Gbps victim to near zero. Husky breaks existing HW and SW isolation (including SR-IOV and vendor QoS); vendors reproduce results and ship firmware fixes.
AI-written summary · about
chapters (9)
- 0:00Intro
- 1:13Existing Solutions for RDMA Performance Isolation
- 1:44Evaluating RDMA Performance Isolation
- 3:20RDMA Microarchitecture Resource Contention
- 5:22Goal - Systematic Test Suite
- 8:15Memory Registration Cause Severe Cache Contention
- 9:32Error Handling Exhaust RX Pipelines
- 10:36Husky: A Test Suite for RDMA Performance Isolation
- 11:05Example - Synthetic Victim
related talks
15mCollie: Finding Performance Anomalies in RDMA Subsystems
01h 24mPreview Session
12mPreview: Networking
15mSecuring RDMA for High-Performance Datacenter Storage Systems
12mCharacterizing and Optimizing Remote Persistent Memory with RDMA and NVM
18mRDMA is Turing complete, we just did not know it yet!
27msRDMA -- Efficient NIC-based Authentication and Encryption for Remote Direct Memory
23mTeRM: Extending RDMA-Attached Memory with SSD
