A SmartNIC-based Load Balancing and Auto Scaling Framework for Middlebox Edge Server
A SmartNIC-based Load Balancing and Auto Scaling Framework for Middlebox Edge Server
复制标题
DOI：
10.1109/nfv-sdn53031.2021.9665167
复制
发表时间：
2021-11
期刊：
影响因子：
--
通讯作者：
Zhen Ni;Cui Wei;Timothy Wood;Nakjung Choi
中科院分区：
文献类型：
--
作者：
Zhen Ni;Cui Wei;Timothy Wood;Nakjung Choi
为蜂窝运营商运行网络功能的边缘云服务器必须同时提供高性能和高资源效率。最近的研究主要集中在大型数据中心的多台服务器之间的负载平衡，但单个主机内的负载平衡却被忽视了。多核服务器利用多队列来将传入的数据包分发到不同的CPU;然而，现有的方法可能会导致不公平的工作分配，因为不同的流可能在大小和处理时间上有很大的差异。虽然NIC具有关于传入流量的详细信息，但它缺乏关于服务器资源的知识，这使得硬件无法完全支持自动扩展和负载平衡功能。对于运行虚拟网络功能的边缘中间盒服务器，这些问题更加严重，其中处理延迟至关重要。SmartVPN提供了一个独特的机会，可以通过增强的可编程性和灵活性来解决此问题。在本文中，我们提出了SmartLB -一个可编程的硬件框架，以探索如何中间盒服务器可以与智能网关一起工作，以提供一个跨层的解决方案。我们的评估表明，SmartLB可以提高流级公平性，减少尾部延迟，并支持高性能网络功能的自动扩展功能。
Edge Cloud servers running network functions for cellular carriers must provide both high performance and high resource efficiency. Recent research has focused on load balancing across multiple servers in a large data center, yet load balancing within a single host has been neglected. Multi-core servers make use of multi-queue NICs in order to distribute incoming packets to different CPUs; however, existing approaches can lead to an unfair distribution of work since different flows may have significant variations in size and processing time. While the NIC has the detailed information about incoming traffic, it lacks knowledge about the server’s resources which prevents the hardware from fully supporting auto scaling and load balancing features. These problems are exacerbated for edge middlebox servers running virtual network functions, where processing latency is critical.SmartNICs offer a unique opportunity to resolve this problem with their increased programmability and flexibility. In this paper we present SmartLB - a programmable hardware framework to explore how middlebox servers can work together with SmartNICs to provide a cross-layer solution. Our evaluation shows that SmartLB can increase flow-level fairness, reduces tail latency, and support auto scaling feature for high performance network functions.