Heavy-Traffic Universality of Redundancy Systems with Data Locality Constraints.

Heterogeneity and compatibility relations between tasks and servers are becoming ubiquitous in cloud computing platforms due to data locality and network topology constraints. These features strongly diverge from the inherent symmetry of the supermarket model as the baseline scenario for performance benchmarking in parallel-server systems. Motivated by that issue, we explore how compatibility constraints impact the system performance in comparison with fully flexible assignment options. We specifically focus on redundancy scheduling systems to gain insight from product-form distributions via a heavy-traffic limit. The asymptotics reveal a striking universality property, in the sense that the system achieves complete resource pooling and exhibits the same behavior across a broad range of scenarios. In particular, the performance of a fully flexible system can asymptotically be matched even with quite stringent compatibility constraints.