Coded Computation Across Shared Heterogeneous Workers With Communication Delay

Distributed computing enables large-scale computation tasks to be processed by multiple workers in parallel. However, the randomness of communication and delays across causes straggler effect, which may degrade delay performance. Coded helps mitigate but amount redundant load task assignment should carefully optimized. In this work, we consider a multi-master heterogeneous-worker distributed scenario, where matrix multiplication are encoded allocated with different capabilities. The goal is minimize plus all tasks. We propose joint worker assignment, resource allocation algorithms under both dedicated fractional policies, each can process from either single master or masters, respectively. Then, non-convex minimization problem solved employing Markov’s inequality-based approximation, Karush-Kuhn-Tucker conditions, successive convex approximation methods. Through extensive simulations, show that proposed reduce completion compared benchmarks.