Intelligent Compute Allocation in Advertising: Value Quantification, Elastic Elimination, and Dynamic Optimization

Background: With rapid growth of ad traffic, orders, and model complexity, compute consumption has become severe. Traffic quality varies, leading to diminishing returns on additional compute resources. Balancing traffic revenue and compute cost is a key challenge.

Since 2022, iQIYI ad engine team has explored intelligent compute allocation to maximize business revenue under compute constraints. Implemented key technologies such as traffic value estimation, elastic elimination, and dynamic compute allocation, achieving significant gains.

How to quantify traffic value : Traffic requests carry ad slot and user features; CTR and conversion predictions are performed to find optimal ad match. Full pipeline evaluation consumes compute, so estimating fair traffic value with minimal compute is challenging.

How to quantify compute : Factors like number of recall channels, bid queue length, model choice affect CPU, memory, latency. Need to convert multiple metrics into a unified compute metric, possibly with weighting.

Optimal value‑compute allocation : After quantifying value (V_i) and compute (C_i), allocate compute to maximize total traffic value under total compute constraint C.

Modeling: Assume linear relationship between compute consumption and traffic value. Parameters k and B are adjusted to keep compute at maximum constraint.

Two‑stage allocation: Stage 1 performs elastic elimination at entry (cut off traffic with V_i < V0). Stage 2 performs fine‑grained compute allocation during recall and ranking.

Value quantification methods :

Online calculation using cache statistics (fill rate, eCPM) with low weight due to sparsity.

Offline statistics on ad slot value over daily windows with high weight.

Model prediction (under exploration).

Compute quantification : Normalize QPS, CPU load, memory usage, latency; use the largest normalized metric as compute indicator. Custom load can be added.

Elastic elimination : Real‑time computation of threshold V0 using PID control to balance traffic value and load. PID formula combines load error (load_c – load_m) to adjust V0. A self‑recovery strategy retains a small proportion of low‑value traffic for potential value increase.

Dynamic compute allocation : Apply DCAF (Dynamic Computation Allocation Framework) to formulate a constrained optimization problem solved by Lagrange multipliers. Decision variables X_ij select compute tier j for traffic i, maximizing expected revenue Q_ij while respecting compute budget q_j.

Results: After deploying the intelligent compute solution, effect‑ad inventory increased >30% with ~5% consumption growth under the same compute scale. System stability improved, achieving >99.99% availability, reducing manual intervention by 3 person‑days per day.

Future work: Extend to global optimization across the whole pipeline, incorporate model predictions and large‑model techniques for more accurate traffic value assessment.