GPU-Accelerated Stochastic Gradient Descent for Scalable Operator Placement in Geo-Distributed Streaming Systems
Modern geo-distributed stream processing systems, particularly those supporting Internet of Things (IoT) workloads, rely on effi cient operator placement strategies to minimize end-to-end latency and avoid overloading resource-constrained edge nodes. Existing approaches, such as NEMO, address this challenge by modeling latency with Euclidean embeddings of network topologies and solv ing operator placement using spring relaxation. However, their CPU-bound optimization process limits scalability, particularly in large topologies with millions of nodes. This paper introduces NEMO-SGD, the first GPU-accelerated, gradient-based optimizer for operator placement in distributed stream processing. NEMO-SGD reformulates the operator place ment problem as adifferentiable loss function and replaces NEMO’s spring relaxation algorithm with a parallelized Stochastic Gradient Descent (SGD) process. Experiments performed on both synthetic and real-world topologies show that NEMO-SGD can optimize placements in under one second for topologies with up to 1 million nodes. This represents a reduction in the optimization time of up to 70% compared to the state-of-the-art NEMO approach. At the same time, NEMO-SGD maintains or even improves the placement quality. Our work shows that gradient-based, GPU-accelerated par allel optimization serves as a practical and scalable foundation for operator placement in next-generation stream processing systems.
