TY - JOUR
T1 - Proactive Resource Autoscaling Scheme Based on SCINet for High-Performance Cloud Computing
AU - Jeong, Byeonghui
AU - Jeon, Jueun
AU - Jeong, Young Sik
N1 - Publisher Copyright:
© 2013 IEEE.
PY - 2023/10/1
Y1 - 2023/10/1
N2 - The container resource autoscaling technique provides scalability to cloud services composed of microservice architecture in a cloud-native computing environment. However, the service efficiency is reduced as the scaling is delayed because dynamic loads occur with various workload patterns. Furthermore, estimating the efficient resource size for the workload is difficult, resulting in resource waste and overload. Therefore, this study proposes high-performance resource management (HiPerRM), which stably and elastically manages container resources to ensure service scalability and efficiency even under rapidly changing dynamic loads. HiPerRM forecasts future workloads using a sample convolutional and interaction network (SCINet) model applied with the reversible instance normalization (RevIN) method. HiPerRM generates a resource request with an elastic size based on the forecasted CPU and memory usage, and then efficiently adjusts the pod's resource request and the number of replicas via HiPerRM's VPA (Hi-VPA) and HiPerRM's HPA (Hi-HPA). As a result of evaluating the performance of HiPerRM, the average resource utilization was improved by approximately 3.96-34.06% compared to conventional autoscaling techniques, even when the resource size was incorrectly estimated for various workloads, and there were relatively fewer overloads.
AB - The container resource autoscaling technique provides scalability to cloud services composed of microservice architecture in a cloud-native computing environment. However, the service efficiency is reduced as the scaling is delayed because dynamic loads occur with various workload patterns. Furthermore, estimating the efficient resource size for the workload is difficult, resulting in resource waste and overload. Therefore, this study proposes high-performance resource management (HiPerRM), which stably and elastically manages container resources to ensure service scalability and efficiency even under rapidly changing dynamic loads. HiPerRM forecasts future workloads using a sample convolutional and interaction network (SCINet) model applied with the reversible instance normalization (RevIN) method. HiPerRM generates a resource request with an elastic size based on the forecasted CPU and memory usage, and then efficiently adjusts the pod's resource request and the number of replicas via HiPerRM's VPA (Hi-VPA) and HiPerRM's HPA (Hi-HPA). As a result of evaluating the performance of HiPerRM, the average resource utilization was improved by approximately 3.96-34.06% compared to conventional autoscaling techniques, even when the resource size was incorrectly estimated for various workloads, and there were relatively fewer overloads.
KW - Cloud computing
KW - container resource autoscaling
KW - resource management
KW - time-series forecasting
UR - http://www.scopus.com/inward/record.url?scp=85164416447&partnerID=8YFLogxK
U2 - 10.1109/TCC.2023.3292378
DO - 10.1109/TCC.2023.3292378
M3 - Article
AN - SCOPUS:85164416447
SN - 2168-7161
VL - 11
SP - 3497
EP - 3509
JO - IEEE Transactions on Cloud Computing
JF - IEEE Transactions on Cloud Computing
IS - 4
ER -