The Pure Dispatcher Pattern: Decoupling Orchestration from Computation in Distributed AI Workloads
Year: 2026
Type: preprint
Abstract: Distributed computing frameworks such as Ray, Spark, and Dask conventionally integrate orchestration and computation within the same execution layer. We argue that this integration is the structural cause of irreducible coordination overhead in embarrassingly parallel workloads with heterogeneous hardware accelerators — overhead that cannot be eliminated by parameter tuning because it is architectural in origin. This paper formalizes the Pure Dispatcher Pattern (PDP): an architectural primitive in which the coordinator is strictly stateless and data-agnostic, performing only work partitioning and result aggregation while never touching computation data. Theorem 1 proves that PDP is coordination-overhead optimal among all correct coordinators via a combinatorial lower bound: any correct coordinator must process at least n work unit identifiers to guarantee completeness, and PDP achieves exactly this bound. Validated on a production corpus of 11,368 pages across 2,246 real-world banking documents, PDP achieves 2.5× improvement over an integrated Ray pipeline at equal resolution and 77.5× reduction in measured coordination overhead. Total speedup with DPI optimization reaches 69.9×.
Author: Jaime, Alejandro
Institution: Universidad Nacional de La Plata
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7155702192
Cites:
Cited by:
Related to: