Prediction of Job Resource Requirements for Deadline Schedulers to Manage High-Level SLAs on the Cloud

Abstract

For a non IT expert to use services in the Cloud is more natural to negotiate the QoS with the provider in terms of service-level metrics --e.g. job deadlines-- instead of resource-level metrics --e.g. CPU MHz. However, current infrastructures only support resource-level metrics --e.g. CPU share and memory allocation-- and there is not a well-known mechanism to translate from service-level metrics to resource-level metrics. Moreover, the lack of precise information regarding the requirements of the services leads to an inefficient resource allocation --usually, providers allocate whole resources to prevent SLA violations. According to this, we propose a novel mechanism to overcome this translation problem using an online prediction system which includes a fast analytical predictor and an adaptive machine learning based predictor. We also show how a deadline scheduler could use these predictions to help providers to make the most of their resources. Our evaluation shows: i) that fast algorithms are able to make predictions with an 11% and 17% of relative error for the CPU and memory respectively; ii) the potential of using accurate predictions in the scheduling compared to simple yet well-known schedulers.