Workload Time Series Cumulative Prediction Mechanism for Cloud Resources Using Neural Machine Translation Technique

Dynamic resource allocation and auto-scaling represent effective solutions for many cloud challenges, such as over-provisioning (i.e., energy-wasting, and Service level Agreement “SLA” violation) and under-provisioning (i.e., Quality of Service “QoS” dropping) of resources. Early workload prediction techniques play an important role in the success of these solutions. Unfortunately, no prediction technique is perfect and suitable enough for most workloads, particularly in cloud environments. Statistical and machine learning techniques may not be appropriate for predicting workloads, due to instability and dependency of cloud resources’ workloads. Although Recurrent Neural Network (RNN) deep learning technique considers these shortcomings, it provides poor results for long-term prediction. On the other hand, Sequence-to-Sequence neural machine translation technique (Seq2Seq) is effectively used for translating long texts. In this paper, workload sequence prediction is treated as a translation problem. Therefore, an Attention Seq2Seq-based technique is proposed for predicting cloud resources’ workloads. To validate the proposed technique, real-world dataset collected from a Google cluster of 11 k machines is used. For improving the performance of the proposed technique, a novel procedure called cumulative-validation is proposed as an alternative procedure to cross-validation. Results show the effectiveness of the proposed technique for predicting workloads of cloud resources in terms of accuracy by 98.1% compared to 91% and 85% for other sequence-based techniques, i.e. Continuous Time Markov Chain based models and Long short-term memory based models, respectively. Also, the proposed cumulative-validation procedure achieves a computational time superiority of 57% less compared to the cross-validation with a slight variation of 0.006 in prediction accuracy.