Quicksand: Harnessing Stranded Datacenter Resources with Granular Computing
summary
Quicksand decomposes applications into small resource proclets—separate compute and memory units—and automatically splits, merges, and migrates them to fill tiny resource gaps. A 10,000‑line prototype on a granular runtime evaluates four applications and reacts in milliseconds to changes; it outperforms a memory disaggregation system by 17% on memory‑imbalanced runs and achieves 2× throughput on CPU‑imbalanced runs while improving balanced throughput by 8% by avoiding excessive threads.
AI-written summary · about
chapters (17)
- 0:00Introduction to Quicksand and stranded resources
- 1:01Stranding problem impact
- 2:46Limitations of granular programming
- 3:49Decoupling via fine-grained units
- 5:12Quicksand design goals
- 6:36Resource proplets compute/memory
- 8:02Quicksand architecture
- 9:05Training preprocessing example
- 10:07Phase 1: Image processing
- 11:30Dynamic compute proplet splitting
- 12:13Phase 2: Memory proplets
- 12:55Deploying to stranded resources
- 13:59Implementation overview
- 14:42Experiment setup
- 15:44Throughput results
- 16:25Fine-grained proplets benefits
- 16:47Conclusion and future
related talks
31mReinvent your cloud strategy: Optimize performance & cut cloud costs
18mNu: Achieving Microsecond-Scale Resource Fungibility with Logical Processes
16mPreview: Cloud Computing
47mCloud Native and Sustainable, Reproducible Scientific Computing
27mFrom Supercomputing to Serving: A Case Study Delivering Cloud Native Foundation Models
18mA Planet-Scale Computer – Abstract Away Regions via Global Service Placer (GSP)
14mMultitenancy for Fast and Programmable Networks in the Cloud
36mQuantum & High-Performance Computing in The Cloud