Research And Optimizations For High Performance I/O Virtualizaion Based On SR-IOV

High performance networking interfaces, such as10Gigabit Ethernet (10GE),are now widely deployed in commercial cloud computing environments. While theplatform is virtualized, the performance of applications running on cloud platformsusually relies on the performance of the I/O performance between virtual machines inthe platform. Therefore highly efcient and scalable I/O virtualization is one of thekey challenge of virtualized cloud computing platforms. Single Root I/O Virtualiza-tion(SR-IOV)isaI/OvirtualizaionstandardreleasedbyPCI-SIGforhighperformancenetwork environments and sR-IOV is becoming more and more popular in virtualizedcloud platforms. SR-IOV eliminates the overhead of redundant data copies and thevirtual network switch through direct I/O to achieve nearly native I/O performance.However, SR-IOV still sufers from serious issues on I/O performance such as redun-dant interrupts and single-threaded network driver.In this paper, we frst study the defects of SR-IOV with10GE networking andindicate two major challenges. Firstly, SR-IOV may generate redundant interrupts un-expectedly when multiple virtual machines are using the I/O device at the same timeand thus result in high CPU overhead. Secondly, SR-IOV also sufers from single-threaded network driver which prevents fully utilizing multi-core computing resourcesand maximizing network bandwidth.We propose two optimizations for enhancing the SR-IOV scalability and perfor-mance. The frst uses adaptive interrupt rate control (AIRC) to reduce CPU overheadcaused by excessive interrupts. We propose a AIRC algorithm based on the mathemat-ical analysis of interrupt rate and I/O performance. We also design a AIRC mecha-nism which maxmizes the throughput as well as minimizes the CPU overhead. When the CPU overhead is reduced, more VMs can be hosted on the same hardware plat-form, thus the scalability is improved. The second is the multi-threaded network driver(MTND) which allows SR-IOV to make full use of multi-core resources. MTND is aparallelized driver model for high performance network such as10GE and MTND canmake the full use of multi-core platform to archieve the maximum network throughput.We implement these optimizations and carry out a detailed performance evalua-tion. The results show that AIRC can reduce CPU overhead by up to51.4%(equivalentto1.4CPU cores saved) as well as retain high throughput and low delay, and MTNDcan improve SR-IOV throughput by up to38%and reduce packet loss rate by up to22%.