Chatty Tenants and the Cloud Network Sharing Problem
summary
Proposes Hadrian, a sharing design that charges each flow by the least-paying partner (taking the min of the two endpoints) and uses VM placement guided by a max‑flow check so per-link weights never exceed capacity. Evaluation includes a 16‑host prototype and 16k‑host simulations, shows inter‑tenant traffic at 10–35%, delivers minimum-bandwidth guarantees while keeping high utilization, bounds a tenant’s maximum network impact to limit abuse, and raises request acceptance by about 20–37%.
AI-written summary · about
chapters (21)
- 0:00Introduction
- 0:34MapReduce Data Analytics
- 1:21Network Sharing Requirements
- 2:54Intertenant Traffic
- 3:42Minimum Bandwidth Guarantee
- 4:33Proportionality
- 5:22Solution
- 6:02Host Motor
- 6:45MultiHost Motor
- 7:29Hierarchical Host Motor
- 8:17Communication Dependence
- 9:11Host compliant bandwidth location
- 10:13Upper bound proportionality
- 10:55Example
- 13:34Results
- 14:28Bandwidth Location
- 15:07Request Acceptance Ratio
- 15:52Conclusion
- 16:24Questions
- 19:25Services
- 19:35Codesign
related talks
27mCloudMirror: Application-Aware Bandwidth Reservations in the Cloud
16mOptimizing Network Provisioning through Cooperation
16mPreview: Cloud Computing
30mThinner Clouds with Preallocation
14mMultitenancy for Fast and Programmable Networks in the Cloud
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
18mCloudcast: High-Throughput, Cost-Aware Overlay Multicast in the Cloud
15mMitigating Scalability Walls of RDMA-based Container Networks