Research And Optimization Of I/O Performance Bottleneck In Embedded Virtualization Environment

With rapid development of Internet on mobile, embedded virtualization technologies is experiencing a period of prosperity. Virtual machine monitor in embedded environment focuses on improvement of performance, and embedded system usually interacts with outside world frequently, as a result of which, I/O virtualization performance is an important factor of embedded virtualization techniques. In the industry of embedded virtualization, I/O virtualization performance must ensure that every guest on platform has the minimal required resources to perform its functions.In this paper, we first introduce embedded virtualization platform KVM/ARM and para-virtualization driver VirtIO. Then we introduce the performance model and conduct benchmarks to reveal current bottleneck lying in the I/O path of VirtIO. During the experiment we find frequent hypercalls contributes most of CPU cycles.Based on previous analysis we propose model to optimize virtualized network I/O performance in embedded environment. Our model borrows an idea of interrupt moderation, instead of invoking the hypercall immediately, our model caches the I/O request for a while and allows one hypercall to serve multiple I/O requests, thus improving CPU utilization and virtualization performance. We conduct experiments and verify the model, after which we introduce design and implementation of key data structure, functions and workflow.Finally we introduce test and verification for the optimization model including experiment platform, method and experiment results. The experiments show that our optimization shows superior performance in terms of throughput. Our optimization improves network performance by 121.14%, 46.23% and 157.42% for bulk data transfer scenario, high concurrency read scenario and combination of read and write scenario respectively. We also evaluate the impact on network latency and the result shows the network latency in our optimization is nearly the same. At last we summary our work, which could shade some light on future work.