Adaptive Resource Provisioning for Virtualized Servers Using Kalman Filters

Abstract

This is the accepted version of the paper.This version of the publication may differ from the final published version. Resource management of virtualized servers in data-centres has become a critical task, since it enables costeffective consolidation of server applications. Resource management is an important and challenging task, especially for multi-tier applications with unpredictable time-varying workloads. Work in resource management using control theory has shown clear benefits of dynamically adjusting resource allocations to match fluctuating workloads. However, little work has been done towards adaptive controllers for unknown workload types. This work presents a new resource management scheme that incorporates the Kalman filter into feedback controllers to dynamically allocate CPU resources to virtual machines hosting server applications. We present a set of controllers that continuously detect and self-adapt to unforeseen workload changes. Furthermore, our most advanced controller also self-configures itself without any a priori information and with a small 4.8% performance penalty in the case of high intensity workload changes. In addition, our controllers are enhanced to deal with multi-tier server applications: by using the pair-wise resource coupling between tiers, they improve server response to large workload increases as compared to controllers with no such resource-coupling mechanism. Our approaches are evaluated and their performance is illustrated on a 3-tier Rubis benchmark web-site deployed on a prototype Xen-virtualized cluster. Permanent repository link

Robust Dynamic CPU Resource Provisioning in Virtualized Servers

“…where x k ∈ [0, 1] is the CPU utilization, i.e., the percentage of the total CPU capacity actually used by the application component during time-interval k. The independent random process w k is the process noise which models the utilization between successive intervals caused by workload changes, e.g., requests being added to or removed from the server; it is often assumed to be normally distributed [7], [10], but it can also be a distribution of finite support [11]. The total CPU utilization of a VM which is actually observed by the Xen Hypervisor, y k ∈ [0, 1], is given by…”

“…The Kalman MIMO controller has been previously evaluated in [10] and in [7]. However, in this experiment the Kalman MIMO controller is compared with the other two MIMO controllers that are designed and implemented in the ViResA project.…”

“…In addition, the resource allocation problem across consolidated virtualized applications under conditions of contention have been considered in [25], [26]: when some applications demand more resources than physically available, then the controllers share the resources among them, while respecting the user-given priorities. Kalyvianaki et al [7], [10] were the first to formulate the CPU allocation problem as a state prediction one and propose adaptive Kalman-based controllers to predict the CPU utilization and maintain the CPU allocation to a user-defined threshold. Even though the standard Kalman filter provides an optimal estimate when the noise is Gaussian, it may perform poorly if the noise characteristics are non-Gaussian.…”

BARISTA: Efficient and Scalable Serverless Serving System for Deep Learning Prediction Services

“…Vertical elasticity adds flexibility as it eliminates the overhead in starting a new VM and loading the service model. Prior efforts to scale the CPU resources vertically appear in [24], [25] including an approach that uses the discrete-time feedback controller leveraging MAPE-K loop for containerized applications [3]. Barista uses an efficient, proactive method to trigger the scaling of resources horizontally while relying on vertical scaling reactively to allocate and de-allocate CPU cores for model correction when our estimation model cannot predict accurately.…”

BrownoutCC: Cascaded Control for Bounding the Response Times of Cloud Applications

“…Computing resource allocation can easily be cast into a control problem, where a controller decides the amount of resource to allocate to different entities based on desired and measurable performance metrics [1,17,21,25]. Recently, the cloud computing domain has emerged as an interesting application domain for control-theoretical principles and techniques [4,7,10,15,22].…”