Fast resource scaling in elastic clusters with an agile method for demand estimation

For energy saving, elastic clusters are introduced to cut back the energy wasted on powering unused servers. In an elastic cluster, the number of working servers, or called resources, is dynamically scaled based on resource demand of workload. However, many traditional scaling methods are unaware of an exact resource demand of workload. They gradually scale resources according to current service level with loose demand estimations or even with no estimation. Additionally, to provide the ability to make precise demand estimations, some other methods are proposed. They artificially represent system situation with a general model, but the model may not well reflect the reality because it is often difficult to describe the real situation of a system. For both of these methods, resources cannot be exactly scaled to the demand when demand changes, and there is a time delay before resources are scaled to the demand. This scaling delay will incur a performance degradation when workload increase, and will cause an energy waste when workload decrease. In this paper, we strive to efficiently estimate the actual demand of workload and achieve fast resource scaling in elastic clusters. Unlike traditional methods which make great efforts to understand the complex system situation, we only concentrate on the information of past actual resource demands. This information is actually the most straightforward and valid reflection to the real situation of a specific system, so it contains valuable knowledge for estimating the actual resource demand of new incoming workload. Therefore, we propose an agile method to directly estimate resource demand based on that knowledge, thus achieving a high accuracy. Specifically, our method directly learns that knowledge through a learning method-random forests, so it does not need artificial system analyses which are both complex and time-consuming. In addition, it is efficient to build random forests and make resource estimations in our method. Thus, our method can be efficiently and agilely performed in elastic clusters to reduce the scaling delay and achieve fast resource scaling.