摘要
Virtualization can provide many benefits for managing resources, including higher resource utilization, lower energy cost, faster fault recovery, and more flexible resource provisioning. However, provisioning resources for applications in the cloud environment has been challenging, especially for scientific applications with more complex runtime behavior and higher performance demand. In this work, we use real scientific applications and performance benchmarking tools to analyze the application performance of our in-house virtualized cluster. We demonstrate that the performance degradation of virtualization can be less than 10% with proper virtual machine configuration and the support of hardware virtualized InfiniBand. Our study on four real scientific applications also proved that the application performance is difficult to model or predict. Therefore, we developed an auto-tuning tool for finding the best resource provisioning setting in terms of both time and cost for any given application. We evaluate our design on an in-house KVM-based virtualized cluster with an InfiniBand connection. Comparing an optimal result from an exhausting search, we verified that our auto-tuning tool achieves accuracy over 90%, comparing to the best deployment, by using much less tuning time and execution runs.