Session-2, Paper-3, NFVnice: Dynamic Backpressure and Scheduling for NFV Service Chains

Authors: Sameer G Kulkarni, Wei Zhang, Jinho Hwang, Shriram Rajagopalan, K. K. Ramakrishnan, Timothy Wood, and Mayutan Arumaithurai and Xiaoming Fu.

Presenter: Sameer G Kulkarni

Presentation: Presenter first motivates by stating that NFV applications are growing and popular and then highlights the problem as they do not scale. These NFs are diverse that require fair scheduling and efficient service chaining. Their diversity stems from their processing and performance requirements. Some middleboxes operate in Mbps speed, while others can only process in Kbps speed. It has been found that these NFs cannot achieve (1) fair scheduling: standard CPU schedulers do not have sufficient information to allocate resources in a way that provides rate-cost proportional fairness. CPU schedulers usually provide fair allocation of processing time, but if computation costs vary between NFs this cannot provide rate-cost fairness. (2) efficient service chaining: Combination of a number of NFs into service chains demands careful resource management across the chain to minimize the impact of bottlenecks. Processing a packet only to have it dropped from a subsequent bottleneck’s queue is wasteful, and a recipe for receive livelock

To address these issues, this paper proposes NFVnice (a user space control framework for scheduling NFV chains). It provides fair and efficient resource allocations to NF service chains. The idea is based on assisted preemptive scheduling, where network functions provide hints to the underlying OS with regard to their utilization. To address efficient service chaining, NFVnice leverages queues between NFs in a service chain to know when NFs in a chain are overloaded or blocked in their operations.

The key components are (1) cgroup: that limits the resources for each user. This paper uses weigh computation algorithm and updates the results every 10ms. (2) Backpressure: that does selective per chain backpressure marking. The proposed mechanism is different from traditional back-pressure in a way that the this paper marks the packet flow (rather individual packets) to drop it at the source.(3) ECN: that signals the congestion, and (4) I/O management library that abstracts implementation complexities from the NF implementation.

The results show that NFVnice can achieve throughput improvement upto 2 times, and avoids CPU wastage through backpressure; and improves resource utilization through scheduling.

No question from audience.