Personalized Computing‑Power Allocation for Alibaba Display Advertising: Transformers Engine and DCAF Algorithm

Background and Motivation Over the past two years, deep‑learning‑driven recommendation systems have saturated, leading to a new bottleneck of limited computing power and diminishing model gains. Alibaba's display‑advertising team has therefore focused on improving computing‑power efficiency to sustain performance growth.

Three‑Stage Power‑Efficiency Evolution The team defines three stages: Power‑Efficiency 1.0 (2018) – point‑wise engineering optimizations; Power‑Efficiency 2.0 (2019) – co‑design of model design and engineering, yielding large gains; Power‑Efficiency 3.0 (2020) – system‑wide, “personalized” power allocation that makes the engine flexible and maximizes business value under a fixed power budget.

Transformers Engine The resulting system, named Transformers (a nod to “Transformers”), integrates power allocation into the engine design, allowing dynamic adjustment of computation per request based on its value‑to‑cost ratio.

Problem Formalization Power allocation is modeled at both micro (per‑request) and macro (time‑varying traffic) levels. For each request i and candidate power tier j, the algorithm knows the expected business value value_{i,j} and power cost cost_{i,j} . The goal is to maximize total value subject to a total power constraint C.

Static Optimal Allocation – DCAF Algorithm The DCAF (Dynamic Computation Allocation Framework) formulates the problem as a constrained optimization with a Lagrangian multiplier λ representing the marginal value‑to‑cost threshold. By solving the dual problem for a given λ, each request independently selects the tier with the highest (value_{i,j}‑value_{i,j'} ) / (cost_{i,j}‑cost_{i,j'}) exceeding λ. λ is tuned via binary search so that total power consumption matches the budget C.

Dynamic Allocation – AllSpark Framework To keep λ optimal as traffic fluctuates, the AllSpark system provides real‑time feedback control. It consists of four components:

Regulation chain : a central Controller collects module‑level metrics (failure rate, latency, CPU usage) and issues average tier suggestions; lightweight Agents embedded in each module pull these suggestions and enforce them.

Feedback chain : metrics are scraped by Prometheus and stored in a time‑series DB, giving the Controller ~5 s latency.

Control panel : UI for configuring targets, strategies, and tier policies per module or control unit, with versioning and gradual rollout.

Monitoring view : visualizes real‑time and historical tier values, alerts on abnormal changes, and ensures observability.

System Flexibility in Practice During the 2020 Double‑Eleven promotion, the AllSpark‑enabled Transformers engine handled a ten‑fold traffic surge with a single baseline deployment, automatically adjusting power tiers to keep latency low and RPM high, delivering millions of RMB incremental revenue. The framework also isolates local failures by throttling power to problematic modules, preserving overall system stability.

Beyond Power Allocation The same co‑design principles extend to multi‑scenario resource sharing (OneEngine), turning the engine into a “brain” that balances algorithmic performance, power usage, and system metrics across the entire online service.

Reference: Jiang et al., “DCAF: A Dynamic Computation Allocation Framework for Online Serving System”, arXiv:2006.09684.