Research Of Adaptive Interrupt Rate Collaboration For High Performance Network Virtualization

The current data centers require high data throughput and real-time data processing. The requirements lead to the development of high performance networks. Virtualization technologies, as a foundation of the data centers, are influenced by the development of high performance networks. Improving the I/O performance of virtual machines under the circumstance of high performance networks is one of the greatest challenges faced by virtualization platforms.Single Root I/O Virtualization(SR-IOV) is a hardware-assisted I/O virtualization solution standardized by PCI-SIG. SR-IOV is promoted due to its applicability in high performance network virtualization environment. With the direct separate access to the I/O devices, SR-IOV diminishes the overhead due to the packet duplication of virtual machine monitor, and achieves I/O performance that is nearly the same performance of physical machines. However, the large amount of packets in high performance network environment causes interrupt processing overhead. The overhead becomes the performance bottleneck of virtual machine network I/O. Interrupt suppression techniques such as NAPI are still insufficient to relieve the burden caused by the massive amount of interrupts.This dissertation first investigates the interrupt rate control of SR-IOV based 10 Gigabit Ethernet NIC, which use the collaboration of NAPI and interrupt throttling. The experimental results demonstrate that the original solution of the NIC is unable to get the best performance in the high load environment, and the main problem is the fixed interrupt rate for different application loads.Aiming at this problem, the dissertation proposes a dynamic interrupt rate collaboration control optimization solution. In order to fully utilize the system resources under different loads and maximize the system throughput, the optimization solution dynamically adjusts the interrupt rate based on the mathematical model analysis of network throughput as well as the statistical analysis of CPU usage. Meanwhile, in order to guarantee the highest response speed under low load, the optimization solution uses the low load control mode of the original control solution and uses an adaptive algorithm for switching mechanisms.The dissertation implements and evaluates the optimization solution. The experimental results demonstrate that the dynamic interrupt rate collaboration control optimization solution is applicable to different application scenarios. In addition, the optimization solution can improve the network throughput by at most 15% while fully utilizing CPU resource and guaranteeing the system response speed.