Performance Isolation in Shared Cloud-based Storage Systems

点击翻译按钮获取中文摘要

Cloud computing represents an exciting new opportunity for businesses to free themselves from building and managing their own computing infrastructure. Business needs in computation and online hosting can be met by simply purchasing resources from Cloud providers. The rate of Cloud adoption, however, has been throttled by the inability of Cloud providers to offer performance guarantees that meet customer requirements. The culprits to this problem are also core tenets of Cloud computing: resource sharing and oversubscription. Cloud providers share each server's resources among multiple Cloud computing services from different businesses; this improves resource utilization and reduces the cost of hosting services. Unfortunately, such sharing can lead to inconsistent and unpredictable performance if there is significant resource contention. The goal of this research program is to facilitate the growth of Cloud computing by improving the state of performance isolation in the Cloud. Strong performance isolation that limits the impact of a service's action on other services competing for the same resources allows Cloud providers to provide consistent and predictable performance. Past work has focused on using virtualization to provide isolation between services contending for resources from a single machine. However, it is often the performance of the underlying Cloud storage systems deployed across thousands of machines that determines the performance characteristics of Cloud-based services. This research program will tackle the performance isolation problem in Cloud storage systems by: identifying the root causes of contention through extensive measurement studies, investigating the impact of different design decisions on performance and contention, exploring innovative solutions to improve performance isolation, and designing and implementing a new Cloud storage system that provides strong performance isolation between services. Overall, by addressing a critical barrier to Cloud adoption through improved performance isolation, this work can have a profound impact on the growing and still volatile Cloud computing industry.