Stochastic scheduling for variation-aware virtual machine placement in a cloud computing CPS

摘要

As the most promising computing paradigm, cloud computing opens up new horizons for the area of high-performance distributed computing. Cyber-physical Systems (CPSs) present novel digital systems, which integrate computation, communication and the control of physical resource. Applied CPSs architecture in cloud computing can provide real-time and scalable resource monitoring and offer time-critical applications. With unrivaled scalability and flexibility, the CPSs based cloud services brings significant convenience to customers in need of elastic computing power. The quality of CPSs based cloud services is, to an large extent, determined by the performance of Virtual Machine (VM) placement algorithm for the data center. VM placement also effect the communication between applications and physical resource distribution in cloud computing CPSs. The traditional VM placement algorithm is built upon the two-tier architecture. With the presence of multi-media applications, the application level controller cannot accurately quantify the varying amount computing resources required by VMs at runtime. Consequently, lacking accurate resource demand for each VM, controller at the data center level cannot generate the VM placement with satisfactory feasibility. This architecture no longer fits the modern data centers. In this paper, the two tier VM placement framework is proposed to resolve this technical challenge. Our LP-based variation-unaware VM placement algorithm generates the VM placement with minimized energy consumption. On the other hand, our feasibility driven stochastic VM placement (FDSP) algorithm works seamlessly with the LP-based algorithm to achieve desirable feasibility of the placement. Our experimental results show that the LP-based variation unaware VM placement algorithm improves the energy consumption by 15.3% on average from the baseline algorithm. For test cases with resource request variations, the FDSP algorithm saves 15.7% energy cost compared to the "worst case scenario" of the traditional VM placement paradigm. On the other hand, it improves the feasibility by 50.0% compared to the "best case scenario".