TY - JOUR
AU1 - Patel, Eva
AU2 - Kushwaha, Dharmender Singh
AB - Recent surge in technology and integration of IoT into Cloud computing has resulted in increasingly heterogeneous workloads with unprecedented compute and storage demands. As Cloud workloads evolve, so also their computational demands, resulting in server loads interspersed by random peaks and troughs and by continuous and periodic demands. A predictive model that accounts for these diverse load patterns can generate more realistic future resource demands which is essential for efficient capacity planning and meeting service level objectives with minimal energy consumption. Long Short-Term Memory (LSTM), the prevalent approach for host load prediction suffers from information decay with long inputs, while hybrid methods using Convolution Neural Network (CNN) and LSTM fail to effectively model different host load patterns. To overcome above limitations, we propose a multistep CPU usage prediction approach named RCP-CL, to model random fluctuations and novel continuous and periodic patterns from contiguous and non-contiguous CPU load values augmented with daily and weekly time patterns, by integrating 1-Dimensional CNN (1D-CNN) and LSTM networks. RCP-CL uses parallel and stacked 1D-CNN layers with kernel size and dilation rates that are guided by the autocorrelation and partial autocorrelation analysis of CPU usage while LSTM derives temporal dependencies from the learnt spatial patterns. Experimental evaluations with Google Trace and Alibaba trace demonstrate impressive learning skills of RCP-CL over the state-of-the-art LSTM and 1D-CNN based host load prediction models, achieving up to 20% improvement in mean squared error for Google trace and up to 22% improvement for Alibaba trace.
TI - An Integrated Deep Learning Prediction Approach for Efficient Modelling of Host Load Patterns in Cloud Computing
JF - Journal of Grid Computing
DO - 10.1007/s10723-022-09639-6
DA - 2023-03-01
UR - https://www.deepdyve.com/lp/springer-journals/an-integrated-deep-learning-prediction-approach-for-efficient-Fb016VzooK
VL - 21
IS - 1
DP - DeepDyve
ER -