Self-adaptive and self-configured CPU resource provisioning for virtualized servers using Kalman filters

Abstract

Data center virtualization allows cost-effective server consolidation which can increase system throughput and reduce power consumption. Resource management of virtualized servers is an important and challenging task, especially when dealing with fluctuating workloads and complex multi-tier server applications. Recent results in control theory-based resource management have shown the potential benefits of adjusting allocations to match changing workloads.This paper presents a new resource management scheme that integrates the Kalman filter into feedback controllers to dynamically allocate CPU resources to virtual machines hosting server applications. The novelty of our approach is the use of the Kalman filter-the optimal filtering technique for state estimation in the sum of squares sense-to track the CPU utilizations and update the allocations accordingly. Our basic controllers continuously detect and self-adapt to unforeseen workload intensity changes.Our more advanced controller self-configures itself to any workload condition without any a priori information. Indicatively, it results in within 4.8% of the performance of workload-aware controllers under high intensity workload changes, and performs equally well under medium intensity traffic. In addition, our controllers are enhanced to deal with multi-tier server applications: by using the pair-wise resource coupling between application components, they provide a 3% on average server performance improvement when facing large unexpected workload increases when compared to controllers with no such resource-coupling mechanism. We evaluate our techniques by controlling a 3-tier Rubis benchmark web site deployed on a prototype Xen-virtualized cluster.

Tackling uncertainty in long-term predictions for host overload and underload detection in cloud computing

“…Thus, an interesting area of future work is combining load burst detection techniques with load prediction techniques to deal with a large variety of cloud load patterns. Second, in addition to the currently used scheme of predicting the next CPU usage value for local resource allocation, more sophisticated schemes based on control theory [32,33], Kalman filters [34] or fuzzy logic [35,36] can be explored. Third, a distributed resource allocation approach should be investigated, where each host agent makes live migration decisions in cooperation with nearby host agents.…”

A control theoretical view of cloud elasticity: taxonomy, survey and challenges

“…(2) Whereas, in the case of vertical elasticity, the objective depends on the nature of reconfigurable resources. For example, there are control solutions, where the control objective is the readjustment of the memory allocation (such as [53,54]) or the readjustment of the CPU allocation (such as [49,[55][56][57][58]). …”

“…In the case of horizontal elasticity, CPU utilisation of the cluster is the common metric used for Reference input [44,50,51,60,61], whereas, in the case of vertical elasticity, the utilisation of individual resources is utilised as Reference input. For example, the control solutions in [55,58] rely on the use of CPU utilisation, whereas Memory consumption was focused in [53] and both components were utilised in [48]. In last, very few approaches including the control solutions proposed in [28,29,75] use both performance based (i.e., Response time) and capacity based (i.e., CPU utilisation) as Reference input.…”

“…The authors of [32,55,62,68,75] focused on vertical elasticity, where they all used the dynamic readjustment of CPU allocation to maintain the target CPU utilisation of the system. Amongst these solutions, Zhu et al [75] used a nested control design that comprise of two integral feedback loops.…”

Elastic Virtual Machine for Fine-Grained Cloud Resource Provisioning

“…Moreover, the work presents scale-up algorithm with no scale-down. To predict next workload and provision enough resources to cope with workload surge, the work in [13] implements two feedback controllers integrated with Kalman filter. The first controller, the Basic Controller (BC), predicts separately the necessary CPU allocation for each tier.…”