Cloud Datacenter Workload Prediction Using Complex-Valued Neural Networks
Kashifuddin Qazi, Igor N. Aizenberg · 2018
Cloud computing infrastructures and datacenters depend on intelligent management of underlying CPU, memory, network, and storage resources. A variety of techniques such as load balancing, load consolidation, and remote memory allocation are used to maintain a fine balance between conflicting goals of high performance, and low costs and energy consumption. To meet these goals, successful prediction of the workloads is an important problem. By accurately predicting the resource utilization of host machines, datacenter owners can better manage the available resources. This paper presents a host resource usage prediction approach, based on a Multilayer Neural Network with Multi-Valued Neurons (MLMVN). An enhancement is further implemented to MLMVN to make it suitable for cloud datacenter applications. The approach is evaluated on real world load traces from Google's cluster data, as well as two grid based load traces. The algorithm is compared against some current state-of-the-art host-load prediction algorithms to show its accuracy, as well as performance gains.