Research Workloads on Preemptible VMs
summary
GCE Preemptible VMs are regular VMs with a 24‑hour limit and 30 seconds termination notice, priced at $0.01 per core hour (~80% cheaper); practical use requires checkpointing and elastic schedulers. Two Sigma adapts Mesos + Cook and a homegrown Launchy autoscaler + Cook to request custom machine types, chase capacity across zones, and absorb ramp‑up latency; they report 5–15% preemption rates, routinely hit huge bursts (hundreds of thousands of cores), achieve ~90–95% allocation but still face hybrid storage, network backpressure, and one‑job tail inefficiencies.
AI-written summary · about
chapters (4)
- 0:00Introduction to preemptible VMs
- 5:59Two Sigma's research workloads
- 28:50Challenges and takeaways
- 31:43Q&A
related talks
33mKubernetes cost optimization - How to actually do it
16mUnlocking unallocated cloud capacity for long, uninterruptible workloads
17mOptimizing Cloud Compute Resources with Spare Cores
30mThinner Clouds with Preallocation
28mAI Workload Preemption in a Multi-Cluster Scheduling System at Bloomberg - Leon Zhou & Wei-Cheng Lai
16mParcae: Proactive, Liveput-Optimized DNN Training on Preemptible Instances
15mExperiences in Managing the Performance and Reliability of a Large-Scale Genomics...
19mStarburst: A Cost-aware Scheduler for Hybrid Cloud