Constant-Factor Scheduling Algorithms for Time-Varying Cloud Capacity

On February 11, 2026, Google Research announced the first constant-factor approximations for non-preemptive throughput maximization when cloud capacity varies over time. The SPAA 2025 paper proves offline greedy and primal-dual bounds, an online restart-based 1/2-competitive algorithm, and a 1/11-competitive common-deadline algorithm. This bridges static scheduling theory with the dynamic reality of data centers.

Summary

Source factOn February 11, 2026, Google Research announced the first constant-factor approximations for non-preemptive throughput maximization when cloud capacity varies over time. The SPAA 2025 paper proves offline greedy and primal-dual bounds, an online restart-based 1/2-competitive algorithm, and a 1/11-competitive common-deadline algorithm.

What happened

Source factGoogle Research published a blog post by Manish Purohit describing new scheduling algorithms for fluctuating capacity. The work was jointly authored with Aniket Murhekar (UIUC), Zoya Svitkina, Erik Vee, and Joshua Wang (Google Research) and presented at SPAA 2025. The post introduces the non-preemptive throughput maximization problem under time-varying capacity and reports several constant-factor guarantees.

Why it matters

AI analysisReal cloud infrastructure experiences constant capacity fluctuations due to high-priority tasks, maintenance, and power limits. For non-preemptive batch jobs, interruptions can waste all work, so schedulers need algorithms that handle uncertain capacity proactively. This research closes a theoretical gap by offering worst-case guarantees, which is valuable for building reliable schedulers in dynamic data centers.

What changed

Source factThe problem setup now includes a time-varying capacity profile and non-preemptive jobs with release times, deadlines, and weights. Offline: greedy achieves 1/2-approximation for unit weights; primal-dual achieves 1/4 for weighted profits. Online with restarts: greedy is 1/2-competitive. Online with discard: all algorithms have competitive ratio approaching zero, except common-deadline instances where a constant 1/11 algorithm exists.

What is actually new

Source factThe authors state they initiate the formal study of throughput maximization under time-varying capacity. They provide the first constant-factor approximation algorithms for offline and online variants. The online restart model's 1/2-ratio matching the offline bound is a new result, as is the 1/11-competitive algorithm for common deadlines and the impossibility result for discard-only online algorithms.

Evidence assessment

AI analysisThe blog post contains high-level proof sketches and claims, but not the full SPAA paper. No experimental data or formal theorem statements are included, so independent verification is not possible from this source alone. The credibility is supported by peer-review at SPAA, but the evidence strength remains moderate for practical adoption.

Constraint shift

AI analysisThe central constraint shift is from static capacity to a time-varying capacity profile. This affects when jobs can be started and whether they should be interrupted. Restart flexibility is shown to make online scheduling tractable, while permanent discard is not. This changes the design space for cloud schedulers, moving from conservative static reservations to dynamic, interrupt-aware policies.

Implications

AI hypothesisCloud providers may experiment with greedy restart-aware schedulers to improve batch throughput under fluctuating capacity, but the worst-case 1/11 guarantee may be too conservative for production. Researchers could explore randomized algorithms or online learning to close the gap between 1/11 and 1/2. Practical impact will likely depend on whether these worst-case bounds translate to average-case performance in real cloud workloads.