Stratus: Clouds with Microarchitectural Resource Management
summary
Stratus proposes treating microarchitectural isolation as a first-class resource with a declarative constraint API (e.g. isolate(llc,1.0,64)), models resources as hard vs soft with a scale 0–1, and shows that isolating L3 can change Redis performance by up to 60%. Allocation uses a cloud knowledge base of topology and live utilization; isolation is paid with isolation credit. Prototype relies on Intel CAT and page coloring for caches, is under development, and must still prove low-latency allocation at 1–10 ms and scale to tens or hundreds of thousands of machines.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:23Once Upon a Time in the Cloud
- 1:04Problems with (Unsupervised) Sharing
- 2:57Stratus: Clouds with Principled Microarchitectural Resource Management
- 3:49Capturing Isolation: A Declarative Interface
- 6:31Building a Reasoning Framework
- 7:43Charging for Isolation: Isolation Credits
- 10:26Summary: Stratus
- 11:26Challenges and Discussion Points
related talks
27mSupporting Security Sensitive Tenants in a Bare-Metal Cloud
30mThinner Clouds with Preallocation
16mPreview: Cloud Computing
21mUpdate on Cloud Foundry Isolation Segments (formerly Elastic Clusters)
18mProgrammable Hyperscale Cloud
15mHere, There and Everywhere: The Past, the Present and the Future of Local Storage in...
37mStay Out of My Yard! Isolation Segments - a New Feature for Isolating Workloads in Cloud Foundry
20mThe Coral Reef Cloud: Autoscaling, Reservation, and Preemption