摘要：In virtual machine system, different workloads are consolidated into a single platform to fully utilize the hardware resources. However, the diversity and strong variation of applications always make it difficult to optimize the resource allocation and thus reduce the system performance and efficiency. Therefore, how to accurately analyze and predict the runtime behavior of applications has become an important basement for virtual machine system optimization. In order to study the characteristic and predictability of virtualization applications, this paper proposes a dynamic behavior characterizing and predicting methodology under Xen virtual machine. We analyze the characteristics of several typical virtualization workloads with fine temporal granularity and apply several online predictors to predict application's runtime I/O behavior. Experiment results demonstrate that the I/O behavior of virtualization workloads can be efficiently predicted by using proper predicting model and configuration. With this result, we further investigate the possibility of virtual machine scheduler optimizing based on I/O behavior characterizing. Several important issues are discussed including I/O computing jobs isolation through asymmetric scheduling, VM dynamic migration based on execution phase tracking and co-scheduling of multiple cooperative virtual machines. Preliminary test results demonstrate that this approach could efficiently reduce the performance degradation caused by scheduling competition in virtual machine system.
