Reduce the Energy Cost of Elastic Clusters by Queueing Workloads with N-1 Queues

Abstract

In Data Centers (DCs), elastic clusters are introduced to cut down the huge energy cost. In elastic clusters, the number of working nodes can be manipulated based on the intensity of workloads. However, affected by the way of distributing workloads to working nodes, the required number of working nodes is different to meet the Service Level Agreement (SLA) of workloads. Workloads consist of several requests which come from clients. In general, workloads are queued and served with N-N queues. The first N means that multiple requests can be queued in the service queue maintained by cluster managers. In addition, the second N means that the service queue of each working node can also queue multiple requests. With N-N queues, requests are first received to the service queue maintained by cluster managers, and then are distributed to appropriate service queues of working nodes. According to queueing theory, a fact is that the service efficiency of N-N queues is lower than that of N-1 queues. Here, N-1 queues mean that the service queue maintained by cluster managers can queue multiple requests, while no request is allowed to be queued in working nodes. Motivated by this fact, we propose an N-1 queueing method to make all service queues work in the form of N-1 queues. Thus under same workloads, fewer working nodes are required to meet a same SLA. As a result, without suffering performance degradation, the energy cost of an elastic cluster can be significantly reduced.