| Recently,with increased user concern over privacy on the Internet,HTTPS is widely deployed by many web services(e.g.video streaming,social network,online shopping).However,the cryptography operations in HTTPS inherently incur costly CPU overhead and infrastructure energy issue.QAT,a commercially available Application Specific Integrated Circuit(ASIC)accelerator,has been widely adopted by many service providers to mitigate the cost and improve HTTPS performance on commodity infrastructure,attributing to its high performance and energy-efficiency.Nowadays,increased numbers of crypto applications are developed on cloud environ-ments,which has led to the demands of utilizing accelerator in the cloud.Unfortunately,because those accelerators are still in lack of scalable and efficient virtualization solu-tions,they are still not fully utilized in the cloud environment.Take Intel~?QuickAssist Technology(QAT)accelerator,one of the most widely deployed accelerators made by Intel,for example.So far,Its SR-IOV solution could only be assigned to fixed VMs which comprises flexibility.And para-virtualization virtio-crypto needs to install a new front end driver in VM,making traditional offloading model unusable.To solve these problems,in this paper,We proposed QVirt:a scalable and efficient para-virtualization solution for SSL/TLS accelerators.By taking advantage of para-virtualization model,accelerators can be virtualized and shared by virtual machines with full features and no number constraints,which can save CPU cycles for other jobs.The main goals of the design are functional equivalence,flexibility,transparency and low overhead.We reuse virtio-crypto,add a new shim layer between application and front-end driver,and leverage user-space dataplane processing framework to achieve these goals.We implement QVirt in Linux KVM/QEMU with QAT.Our evaluation shows that QVirt is capable of providing transparent cryptography service to applications.And virtual crypto device provides most features that QAT can provide.QVirt outperforms traditional software solution by a factor of 4.8,and we believe there is still a lot of room for improvement. |