Structure-Aware Serverless Orchestration via Graph Reinforcement Learning: Joint Optimization of Function Fusion and Placement

Abstract

Function-as-a-Service (FaaS) platforms are increasingly used for data-intensive applications that execute as Directed Acyclic Graphs (DAGs). In these workflows, function fusion and placement are coupled: fusing dependent functions can reduce data transfer and cold-start overhead, but it also changes the memory footprint and the set of feasible deployment nodes. We propose GRL-Orch, a graph-reinforcement-learning orchestration framework for this coupled control problem. GRL-Orch encodes workflow topology with a Graph Neural Network (GNN), uses a factorized joint policy for fusion and placement, and applies feasibility-aware repair and placement masks before residual Out-Of-Memory (OOM) penalties are considered. Trace-derived simulator results using Azure Functions 2019 profiles show that the learned GRL-Orch policy reaches a mean objective of 1.147 on 100 held-out workflows, below the strongest Critical-Path-Fusion baseline at 1.184. An extended 500-workflow audit preserves the advantage, with objectives 1.368 versus 1.449. A live Apache OpenWhisk replay on small DAGs further shows consistent fusion-policy ranking in a 15-setting core matrix, while a 24-setting stress extension preserves 20/24 best-policy agreement and exposes deviations on size-20 diamond workflows. Together, these results characterize GRL-Orch as a structure-aware control design for serverless workflow orchestration, combining trace-derived scale with OpenWhisk grounding for key fusion-policy trends.