Managing Traffic Demand Uncertainty in Replica Server Placement with Robust Optimization

Abstract

Abstract. The replica server placement problem determines the optimal location where replicated servers should be placed in content distribution networks, in order to optimize network performance. The estimated traffic demand is fundamental input to this problem and its accuracy is essential for the target performance to be achieved. However, deriving accurate traffic demands is far from trivial and uncertainty makes the target performance hard to predict. We argue that it is often inappropriate to optimize the performance for only a particular set of traffic demands that is assumed accurate. In this paper, we propose a scenario-based robust optimization approach to address the replica server placement problem under traffic demand uncertainty. The objective is to minimize the total distribution cost across a variety of traffic demand scenarios while minimizing the performance deviation from the optimal solution. Empirical results demonstrate that robust optimization for replica server placement can achieve good performance under all the traffic demand scenarios while nonrobust approaches perform significantly worse. This approach allows content distribution providers to provision better and predictable quality of service for their customers by reducing the impact of inaccuracy in traffic demand estimation on the replica server placement optimization.

A Survey on Replica Server Placement Algorithms for Content Delivery Networks

“…Optimal algorithm such as branch and bound is used to solve the ILP. The simulation study [86] shows that the scenario-based replica server placement outperforms the replica server placement approach with deterministic traffic demand. All replica server placement algorithms in the "Others" category are summarized in TABLE V.…”

“…Ho et al [86] present a replica server placement model that is robust to variations in traffic volumes. Uncertainty in traffic demand has a significant impact on CDN performance as different scenarios result in different traffic volumes in CDNs.…”

“…As a result, with a given placement of replica servers, CDN providers may have to deal with high cost in delivering content to end-users in worst-case traffic scenarios, such as flash crowds. Ho et al [86] present a robust replica server placement model that minimizes the delivery cost across a variety of traffic demand scenarios and minimize the deviations from the optimal solution for each scenario. For a single scenario, the model is reduced to uncapacitated Kmedian model.…”

“…It is mostly because majority of the algorithms adopt the closest server policy and hence the end-users load is not split among the replica servers. Apart from Ho et al [86], Nguyen et al [54] all other algorithms for traditional CDNs lack the resiliency against demand uncertainty, leading to downgraded QoS during sudden spike in demand. In terms of placement granularity, none of the algorithms has high granularity i.e.…”

Design of Robust and Efficient Edge Server Placement and Server Scheduling Policies: Extended Version

“…One natural idea to deal with workload uncertainty is to use robust optimization. This idea has been used to study the service scheduling problem [20] in EC and the replica server placement problem [21] in CDN. However, robust optimization cannot offer any guarantee when the workload is outside of the predicted set.…”