Dynamic Control of Data Streaming and Processing in a Virtualized Environment
Junwei Cao,Wen Zhang,Wei Tan
DOI: https://doi.org/10.1109/tase.2011.2182049
IF: 6.636
2012-01-01
IEEE Transactions on Automation Science and Engineering
Abstract:Performance of data streaming applications is co-determined by both networking and computing resources, and therefore they should be co-scheduled and co-allocated in an integrated and coordinated way. Dynamic control of resource scheduling and allocation is required, because unilateral redundancy in either networking or computing resources may result in the overprovision of it and the other may become a bottleneck. To avoid resource shortage as well as overprovision, in this paper, a virtualized platform is utilized to implement data streaming and processing. In this platform, fuzzy logic controllers are designed to allocate CPU resources; iterative bandwidth allocation is applied and is processing-and storage-aware to guarantee on-demand data provisioning. Experimental results show that our approach leads to higher application performance as well as higher resource utilization, compared with other resource scheduling and allocation methods.Note to Practitioners-Data streaming has become an important paradigm in many business and scientific applications, such as fraud detection in banking industry and gravitational-wave observation in astronomy. In this paper, we apply virtualization to provide better support for these applications, leveraging its advantage in resource on-demand allocation. A novel dynamic control method is proposed so that CPU and bandwidth can be co-scheduled and co-allocated since for data streaming applications these resources are tightly coupled from the performance perspective. In this control method, fuzzy logic control is applied for CPU allocation and an iterative algorithm is adopted for processing-, congestion-, and storage-aware bandwidth allocation. We use fuzzy control because it does not rely on the mathematical modeling of an object and can implement human experts' heuristic knowledge via IF-THEN rules. Therefore, it is desirable in our data streaming scenario, due to the variable coupling and heavily nonlinear nature of the system.