New generative Markov model tames distributed computing complexity

Emerging distributed computing paradigms like the computing continuum are inherently heterogeneous, stochastic, and complex, making efficient resource utilization a challenge. To address this, Alfreds Lapkovskis, Ali Beikmohammadi, Sindri Magnússon, and Praveen Kumar Donta propose a general framework that models distributed systems as a generative Markov model, factorized over a structured system state. The state decomposes into high-dimensional variables, each further factorized over its elements to reflect sparse dependency structures inherent in distributed computing. This factorization yields a tractable model capable of simulation, inference, and policy learning—bridging distributed computing with Markov chain theory and reinforcement learning (RL).

In a case study on collaborative AI inference, a dedicated server combines its resources with volunteered user devices. The model reveals that centralized scheduling becomes a bottleneck at scale, while distributing computation across user devices significantly reduces both latency and server resource consumption. These findings highlight the value of adaptive decision-making in distributed computing systems and validate the framework's utility for modeling, simulation, and optimization. The work is accepted at the 40th International Symposium on Distributed Computing (DISC 2026).