Scheduling in a changing world: Maximizing throughput with time-varying capacity

2026-02-11 · Source: The latest research from Google · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Expert, medium

Summary

Google Research introduced new, provably effective algorithms for non-preemptive job scheduling on cloud infrastructure with time-varying machine availability, a common scenario in large-scale cloud computing. The research, presented at SPAA 2025, addresses the problem of maximizing throughput (total weight or number of successful jobs) when resources fluctuate due to hardware failure, maintenance, or high-priority tasks claiming capacity. For the offline setting, where future capacity is known, a Greedy strategy achieved a 1/2-approximation for unit profit jobs and a 1/4-approximation using a primal-dual framework for varying profits. In the more complex online setting, where jobs arrive dynamically, standard non-preemptive algorithms fail. The team developed a variant of Greedy for "interruption with restarts" that also achieves a 1/2-competitive ratio. For the stricter "interruption without restarts" model, a novel algorithm achieved a 1/11 competitive ratio for common deadline instances, representing the first constant competitive ratio for this problem.

Key takeaway

For AI Scientists designing or optimizing cloud resource schedulers, this research highlights the necessity of moving beyond static capacity assumptions. You should evaluate your current scheduling algorithms against the challenges of time-varying capacity, particularly for non-preemptive batch jobs. Consider implementing algorithms that allow for job restarts to significantly improve throughput in dynamic online environments, or explore the proposed constant competitive ratio algorithms for stricter "interruption without restarts" scenarios, especially when jobs share common deadlines.

Key insights

New algorithms optimize non-preemptive job scheduling in dynamic cloud environments with fluctuating machine capacity.

Principles

• Greedy strategies can achieve strong approximation guarantees.
• Flexibility in job interruption significantly improves online scheduling.
• Constant-factor approximations are crucial for NP-hard problems.

Method

The online algorithm for "interruption without restarts" maintains a tentative schedule, modifying it by adding, replacing, interrupting, or discarding jobs based on arrival and current state, particularly for common deadline scenarios.

In practice

• Prioritize jobs with common deadlines for online scheduling.
• Consider job restart flexibility for better throughput.
• Implement Greedy-based heuristics for offline planning.

Topics

• Time-Varying Capacity Scheduling
• Non-Preemptive Job Scheduling
• Throughput Maximization
• Approximation Algorithms
• Online Algorithms

Best for: AI Scientist, AI Researcher, Research Scientist, AI Engineer

Related on AIssential

• Online LLM Selection via Constrained Bandits with Time-Varying Demand — Online LLM selection under dynamic demand and hard/soft constraints can be optimized using constrained bandit learning.
• Optimal Scheduling in a Question-Answering Forum of Knowledge Workers — The paper models optimal scheduling for paid expert QA forums to maximize stable request handling capacity.
• DataCenterGym: A Physics-Grounded Simulator for Multi-Objective Data Center Scheduling — DataCenterGym simulates geo-distributed data center scheduling, integrating thermal, power, and workload dynamics for comprehensive policy evaluation.
• Petri Net Induced Heuristic Search for Resource Constrained Scheduling — RCPSP can be optimally solved via A* search on a TTPNR reachability graph using consistent combined heuristics.
• Managing Task Execution for Unknown Workloads in Batteryless IoT: A Hardware-Agnostic Evaluation — Dynamic, hardware-agnostic scheduling methods improve reliability for batteryless IoT by adapting to unknown workloads and volatile energy.
• Optimizing cloud economics with linear elastic caching — Dynamic, cost-aware cache sizing using a "ski rental" model significantly reduces total ownership cost by optimizing memory use.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The latest research from Google.