ResQ: Enabling SLOs in Network Function Virtualization
summary
ResQ enforces SLOs by combining Intel RDT cache allocation (CAT) with careful I/O buffer sizing to fix a discovered leaky DMA problem that lets DMA pollute the LLC. Experiments show direct isolation cuts max throughput degradation from 44.3% to 2.9% and 95th‑percentile latency degradation from 24.5% to 2.5%. Admission uses fast profiling and a greedy placer (with a background MILP reoptimizer), and cluster evaluations report 60%–236% better resource efficiency versus prior prediction‑based schemes.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:28NFV Builds on Resource Sharing
- 1:18Offering Performance Guarantees Is Challenging
- 2:01Assumptions on Resource Sharing and Isolation
- 3:26Does Resource Contention Matter?
- 5:56Approaches to Offer Performance SLOS
- 8:44Enabler: Hardware Resource Isolation
- 10:38LLC Isolation Is Not Sufficient!
- 11:55The Leaky DMA Problem
- 13:51Accuracy of ResQ's Isolation Mechanism
- 15:21ResQ Admission Process
- 16:48ResQ Optimal Scheduler
- 17:49Resource Efficiency
related talks
23mNetwork SLOs: Knowing When The Network Is The Barrier To Application Quality Of Experience
21mReexamining Direct Cache Access to Optimize I/O Intensive Applications for Multi
14mUnderstanding RDMA Microarchitecture Resources for Performance Isolation
25mResilience for AI Workloads at Scale: The Fast and the Finicky!
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
01h 24mPreview Session
16mFLB: Fine-grained Load Balancing for Lossless Datacenter Networks
14mODRP: On-Demand Remote Paging with Programmable RDMA