SOPRAN: Integrative Workload Modeling and Proactive Reoptimization for Virtual Machine Management
Jian Hong Zhou, Lei Shi, Kian‐Lee Tan · INTERNATIONAL JOURNAL OF NEXT-GENERATION COMPUTING · 2011
For a data center to operate effectively (i.e., meeting customers’ Service Level Agreements (SLAs)) and efficiently (i.e., minimizing resource consumption), the virtual machines (VMs) must be carefully managed. In particular, as the resource demands of VMs change, the assignment of VMs to physical machines becomes sub-optimal. VM replication and migration provide a solution for dealing with dynamic workloads. However, as migrations are costly, an effective control policy is critical to avoid frequent migrations. Moreover, an agile decision making component is also important to reduce feedback latencies. In this paper we propose SOPRAN, a virtual machine management system leveraging an integrative workload model for the data center, that can dynamically adapt the assignment of VMs to physical machines to minimize resource consumption without sacrificing the SLAs. Different from existing trace-based methods for this problem, SOPRAN characterizes the dynamic workloads in the system using an integrative risk cube model, and approximates the workload demands with a representative state set. The optimal plan for each representative state is incrementally generated, forming the switchable plan set. At runtime, a two-phase re-optimization strategy matches the current system demand to the closest representative state and actuates the corresponding plan in the switchable plan set. At the same time, online monitors profile the actual demands and refine the risk cube to guarantee the model’s accuracy. This modeling technique and optimization procedure based on it brings the great savings in optimization cost and migration opportunities, and enables the high scalability of SOPRAN. We evaluated SOPRAN against the state-of-the-art IBM MFR algorithm. The results show that, with comparable resource consumptions, SOPRAN can achieve more stable SLA violation rate of no more than 4%, 80% lower migration rate, and save up to 90% reoptimization overhead.