Virtualizing Traffic Shapers for Practical Resource Allocation
Gautam Kumar, Srikanth Kandula, Peter Bodík, Ishai Menache · IEEE International Conference on Cloud Computing Technology and Science · 2013
Many network resource allocation scenarios would benefit from the use of traffic shapers, such as weighted fair queues (WFQs) [2], priority queues [10] and rate limiters [11]. However, the number of such shapers implemented in hardware in switches, routers, and network interface cards (NICs) is very low; typically less than ten (see Table 1). Instead of simple use of hardware traffic shapers, network operators thus have to resort to more complex solutions. For example, public clouds such as Amazon Web Services and Windows Azure want to limit the network bandwidth that is allocated to each VM along each path through their network. However, NICs only support a small number of rate limiters in hardware [11]. So, the cloud providers implement rate limits in software, forcing all server traffic to go through the hypervisor. This results in both worse latency and lower throughput because it requires more copying (between guest VM and hypervisor) and it disallows multi-core optimizations such as SRIOV and Direct-IO which let VMs directly read and write from NIC buffers. Dedicating more cores to the hypervisor improves throughput but the cloud provider now has fewer to allocate to customers. As another example, major companies have announced centralized traffic engineering using SDN switches [6]. To drive networks at high utilization, these schemes solve a global optimization problem to map traffic across all the network paths and allocate available network bandwidth among hosts based on business requirements. Here too, there is a need to ensure that hosts adhere to their allocated rates along each path. Further, there is a need to prioritize among traffic classes; e.g., customer-facing traffic where extra latency means lost revenue should be prioritized over bulk data transfers that only care about finishing within a deadline. However, the number of traffic classes is several orders of magnitude larger than the number of available shapers in modData Path Control