Intelligent Compute Allocation for Meituan Takeaway Advertising: Design, Implementation, and Experiments

This article presents the concept of "intelligent compute"—fine‑grained, value‑driven allocation of computing resources—to improve efficiency and reduce cost in Meituan Takeaway's advertising system.

Business background : Takeaway orders exceed 40 million per day, and ad traffic shows a double‑peak pattern (lunch and dinner). During peaks the system faces high load, while off‑peak periods suffer from under‑utilized resources, leading to low overall compute efficiency.

Overall idea : Under a global compute‑capacity constraint, allocate different amounts of compute to traffic of different business value. The solution consists of four key elements: (1) traffic‑value quantification, (2) traffic‑compute quantification, (3) system‑compute capacity quantification, and (4) intelligent compute allocation via "elastic actions" (elastic queue, model, channel, and link) and corresponding "elastic tiers".

Challenge analysis : The main challenges are (a) solving the optimization problem of maximizing traffic value under compute constraints, (b) ensuring system stability when shifting from equal to dynamic compute distribution, and (c) maintaining generality and extensibility across multiple ad‑related scenarios.

Solution design : A multi‑action decision framework is built, comprising decision, collection, and control components. The core decision module provides optimal tier selection and system‑stability guarantees. Elastic actions are defined as follows: elastic queue (different candidate‑list lengths), elastic model (different model sizes), elastic channel (different recall channels), and elastic link (different pipeline complexities). These actions map to discrete tiers that the optimizer selects.

Optimal tier decision : Offline, historical traffic is replayed and a binary‑search algorithm finds the optimal λ that balances value and compute. Online, each request evaluates candidate tiers and selects the one with maximal expected benefit, applying the corresponding elastic actions.

System stability guarantee : Standard mechanisms (circuit‑break, degradation) are combined with a PID‑based real‑time control loop that monitors CPU/GPU utilization, QPS, latency, and failure rate, adjusting compute allocation parameters to keep the system within target thresholds.

Experiments : Two experiments were conducted. (1) With equal machine resources, business revenue increased by 2.3 % (CPM uplift). (2) With equal revenue, machine resources were reduced by ~40 % while maintaining revenue. Both experiments demonstrated that intelligent compute can shift resources from low‑value to high‑value traffic, improving overall efficiency.

Conclusion & outlook : The intelligent compute framework successfully reduces waste and boosts revenue in a large‑scale ad serving system. Future work includes exploring evolutionary algorithms and reinforcement learning for end‑to‑end tier optimization, and tighter integration with elastic scaling infrastructure.