Thinner Clouds with Preallocation

on YouTube ↗30m

" summary

Proposes preallocation: images for multiple services are loaded and paused on a PM so only one VM runs at a time, cutting reserve waste. Uses a greedy two-phase algorithm that runs every 5 minutes with a 10-minute boot cost, simulates 50 services with flash‑crowd traces (30min ramp/steady/drop, spikes from mean 4 to 80), targets 1% overflow, and reports a 35–50% reduction in required physical machines versus classic slack allocation.

AI-written summary · about

" chapters (7)

1. 0:00Introduction
2. 4:46Load
3. 6:58Preallocation
4. 13:18Greedy
5. 17:34Practical considerations
6. 19:17Algorithms
7. 20:28Question

" related talks

• 16mPreview: Cloud ComputingKshiteej Mahajan · OSDI '21 / USENIX ATC '21
• 16mOptimizing Network Provisioning through CooperationRanjita Bhagwan, Yogesh Bansal, Harsha Sharma, Parth Thakkar, Sagar Bharadwaj, Venkata N. Padmanabhan, Vijay Kumar, Kathleen Voelbel · 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22)
• 30mCapacity Management for Fun & ProfitAly Fulton · SREcon21
• 27mSupporting Security Sensitive Tenants in a Bare-Metal CloudMania Abdi, Trammell Hudson, Amin Mosayyebzadeh, Apoorve Mohan, Sahil Tikale, Nabil Schear, Charles Munson, Larry Rudolph, Gene Cooperman, Peter Desnoyers, Orran Krieger · USENIX ATC '19
• 21mChatty Tenants and the Cloud Network Sharing ProblemChanghoon Kim, Hitesh Ballani, Keon Jang, Thomas Karagiannis, Dinan Gunawardena, Greg O'Shea · NSDI 2013
• 32mResearch Workloads on Preemptible VMsMichael Basilyan · Google Cloud Next '18
• 15mStratus: Clouds with Microarchitectural Resource ManagementKaveh Razavi, Animesh Trivedi · HotCloud '20
• 36mSustainable cloud infrastructure: from ideation to deploymentAlban Boitier, Arnaud Kalck · apidays Paris 2020