A Survey of Fast Packet I/O Technologies for Network Function Virtualization

Abstract

Network Function Virtualization (NFV) aims at bringing the benefits of virtualization to network middleboxes (routers, firewalls, Intrusion Detection Systems, ...). In the last few years the NFV use-case, initially hampered by the poor performance of traditional virtualized-I/O and network stacks, has prompted the design of several frameworks, all trying to provide a fast network for VMs and/or containers. These solutions share many common ideas, but also differ in performance, flexibility, portability, amount of specialized hardware required and/or software to be rewritten, attention to energy consumption issues, and so on. In this survey we focus on the NFV data-path, as opposed to the orthogonal control-path. We define a set of desirable features for NFV data-paths and use them to compare a selection of the most promising and/or widely used NFV frameworks. No single solution is optimal for all the features, so our survey may prompt for further research in this area

On the Use of Kernel Bypass Mechanisms for High-Performance Inter-container Communications

“…Another survey among common networking set-ups for high-performance NFV exists [8], accompanied by a quantitative comparison addressing throughput and CPU utilization of SR-IOV, Snabb, OVS with DPDK and Netmap [16]. Authors highlight how the different solutions have remarkable differences in security and usability, and they show that, for local VM to VM communications, Netmap is capable of reaching up to 27 Mpps (when running on a 4GHz CPU), overcoming SR-IOV due to its limited internal switch bandwidth that becomes a bottleneck.…”

Survey of Performance Acceleration Techniques for Network Function Virtualization

“…Due to the different numbers of intermediate abstraction layers, implementation details, and the heterogeneity of possible use cases, there is no single hypervisor that performs best in all scenarios and therefore should be chosen carefully to match the particular requirements [16]. Furthermore, high-speed I/O virtualization plays a crucial role for meeting performance requirements and maximizing flexibility when VNFs are run inside VMs [17]. To this end, drivers such as virtio [18] and ptnet [19] provide a common device model for accessing virtual interfaces.…”

Hardware Acceleration for Container Migration on Resource-Constrained Platforms

“…That is, overall, the use of an accelerator device requires eight copies compared to four copies for a conventional SW implementation. Although data copying from the kernel space memory to the user space memory can be avoided using advanced data processing techniques, such as kernel bypass [48], [49], shared virtual memory using Process Address Space ID (PASID) [50], and Poll Mode Drivers (PMD) (e.g., Data Plane Development Kit (DPDK)), these mechanisms require application and OS/hypervisor kernel modifications which are typically not considered in the networking stack on resource-constrained client devices. Therefore, our evaluations are based on standard techniques and mechanisms that are typically used in resource-constrained client devices.…”