A METHOD OF DYNAMIC ASSIGNMENT OF NEURAL INFERENCE SUBTASKS IN HETEROGENEOUS DISTRIBUTED MOVING-OBJECT RECOGNITION SYSTEMS

This paper addresses the assignment of neural inference subtasks to the computing nodes of a heterogeneous distributed moving-object recognition system. Heterogeneity is shown to be not a secondary complication but an independent factor that qualitatively changes system behaviour: under uniform assignment the sustainable throughput bound is governed by the slowest node, whereas under performance-proportional assignment it equals the sum of the service rates of all nodes. Both bounds are derived from the stability condition of a single-server queue, and the latter is proved to be an upper bound for any assignment policy. A heterogeneity coefficient is introduced as the relative throughput loss caused by uniform assignment, and a non-obvious property of it is established: it grows with the number of nodes at a fixed performance spread, meaning that the cost of a wrong assignment grows with the scale of the system. A dynamic assignment method based on a multi-criteria cost function is proposed, whose terms correspond to the computational, queueing and network components of latency; it is proved that under saturation greedy assignment driven by this function distributes the load proportionally to the service rates of the nodes and therefore attains the established upper bound. The assignment complexity is O(N·M) versus O(N³) for the Kuhn–Munkres algorithm, which for a queue of 800 subtasks yields 1.50 ms against 20.55 ms. Discrete-event simulation confirms that the observed latency knee points coincide with the analytical bounds to within 1 frame per second, and that the proposed assignment outperforms cyclic assignment under overload and under link degradation.