Resource allocation in large-scale multi-server systems

The focus of this dissertation is the task of resource allocation in multi- server systems arising from two applications – multi-channel wireless com- munication networks and large-scale content delivery networks. The unifying theme behind all the problems studied in this dissertation is the large-scale nature of the underlying networks, which necessitate the design of algorithms which are simple/greedy and therefore scalable, and yet, have good perfor- mance guarantees. For the multi-channel multi-hop wireless communication networks we consider, the goal is to design scalable routing and scheduling policies which stabilize the system and perform well from a queue-length and end-to-end delay perspective. We first focus on relay assisted downlink networks where it is well understood that the BackPressure algorithm is stabilizing, but, its delay performance can be poor. We propose an alternative algorithm - an iterative MaxWeight algorithm and show that it stabilizes the system and outperforms the BackPressure algorithm. Next, we focus on wireless networks which serve mobile users via a wide-area base-station and multiple densely deployed short- range access nodes (e.g., small cells). We show that traditional algorithms that forward each packet at most once, either to a single access node or a mobile user, do not have good delay performance and propose an algorithm (a distributed scheduler - DIST) and show that it can stabilize the system and performs well from a queue-length/delay perspective. In content delivery networks, each arriving job can only be served by servers storing the requested content piece. Motivated by this, we consider two settings. In the first setting, each job, on arrival, reveals a deadline and a subset of servers that can serve it and the goal is to maximize the fraction of jobs that are served before their deadlines. We propose an online load balanc- ing algorithm which uses correlated randomness and prove its optimality. In the second setting, we study content placement in a content delivery network where a large number of servers, serve a correspondingly large volume of con- tent requests arriving according to an unknown stochastic process. The main takeaway from our results for this setting is that separating the estimation of demands and the subsequent use of the estimations to design optimal content placement policies (learn-and-optimize approach) is suboptimal. In addition, we study two simple adaptive content replication policies and show that they outperform all learning-based static storage policies.