Improving Load Balancing for Modern Data Centers Through Resource Equivalence Classes

Load balancing is one of the most significant concerns for data center (DC) management, and the basic method is reassigning applications from overloaded servers to underloaded servers. However, to ensure the service availability, during the reassignment of an application, some resources (i.e., transient resources) are consumed simultaneously on its initial server and its target server, which imposes a challenge for load balancing. The latest research has proposed a concept called resource equivalence class (REC: a set of resource configurations such that a latency-critical (LC) application running with any one of them can meet the QoS target). In this paper, we use the REC to improve the load balancing for a DC where multiple LC applications have already been co-located on servers with the service availability and QoS requirements. We formulate the proposed load rebalancing problem as a multi-objective constrained programming model. To solve the proposed problem, we propose to use a machine learning-based classification model to construct the RECs for applications, and we develop a local search (LS) algorithm to approximate the optimal solution. We evaluate the proposed algorithm via simulated experiments using real LC applications. To our knowledge, it is the first time to use REC for improving load balancing.