Abstract
Server consolidation in cloud environment largely enhances the resource utilization, and it also become a good solution to save energy and cost. Virtualization technology is a key technique to realize server consolidation. However, virtualization and server consolidation may cause some problems on system performance, and there are still not any systematic methods or tools for analyzing them. This makes it pretty difficult to predict performances for a cloud environment and to implement a performance guaranteed dynamic provisioning system of computing resources. By the way, there are often heterogeneities of physical resources existed in cloud environments, and they largely increase the complexities of performance prediction and management. In this paper, we propose a suite of fast automated benchmarking tools for cloud platforms, which can be used to evaluate how virtualization and server consolidation effect the performances of applications running on cloud platforms. Its main functionalities are analyze the performance variations of CPU, memory, and disk I/O for virtualization platforms. We can more precisely see the goods and bad of a cloud platform by evaluating those resources separately. Furthermore, for enhancing the efficiency of those performance evaluations, the tools parallelize each performance evaluation by utilizing the property of cloud platforms. Comparing with traditional benchmarking software, the tools can systematically, precisely, and quickly report how virtualization and server consolidation effects the performances of applications. We conduct the fastest and most complicated experiments on several of cloud servers and compare the differences among their results. According to the results of most complicated experiments, we analyze the performances of this cloud platform. In profiling disk I/O performances, we also conduct the experiments and analyze its performances under different image formats and SSD. Finally, we redo the most complicated experiments by parallel profiling. To verify the precision of parallel profiling, we compare its results with original results. According to the experiment results, the results of the fastest experiments can precisely show the most general performances for a cloud platform, and that of parallel profiling also achieve the expected precision. Besides, we also analyze the performance tendencies and virtualization overheads for the workloads of each resource type.