EyeQ: Practical Network Performance Isolation at the Edge
summary
EyeQ enforces per-VM bandwidth guarantees at the server edge using per-VM virtual links and explicit rate feedback (RCP) sampled from incoming packets, implemented as a 10 Gb/s kernel module and a Windows filter driver. A 16-node eval runs 144,000 memcached set/sec (6 KB objects) while an adversarial UDP tenant blasts bursts; EyeQ keeps memcached’s 99.9th percentile latency close to dedicated performance and converges in milliseconds.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:21The past
- 0:45The cloud
- 0:55Performance variability
- 2:10Current status quo
- 3:05The problem
- 3:56Condition study
- 4:33Measurement study
- 6:07EyeQ
- 7:23Example
- 11:55Time Scales
- 12:48Failure Scenarios
- 14:14EyeQ Overview
- 15:03Rate Limiters
- 16:09Packet Rate Test
- 16:47Latency Test
- 19:15Baseline Performance
- 20:47Questions
related talks
16mAn edge-queued datagram service for all datacenter traffic
21mFine-Grained Isolation for Scalable, Dynamic, Multi-tenant Edge Clouds
20mNetKernel: Making Network Stack Part of the Virtualized Infrastructure
22mLIVE Q&A #3
16mBuilding A CSFQ-Inspired Transport for Switched CXL Memory Pooling
16mQueue-Mem: Energy-Efficient Hardware Storage for Advanced Network Function Acceleration
31mLive Q+A 1
25mResQ: Enabling SLOs in Network Function Virtualization
