Stringer: Balancing Latency and Resource Usage in Service Function Chain Provisioning

Abstract

Network Functions Virtualization (NFV) enables telecommunications infrastructure providers to replace specialpurpose networking equipment with commodity servers running virtualized network functions (VNFs). A provider utilizing NFV faces the Service Function Chain (SFC) provisioning problem of assigning VNF instances to nodes in the physical infrastructure (e.g. datacenters), and routing Service Function Chains (sequences of functions required by customers, a.k.a. SFCs) in the physical network. The provider must balance competing goals of performance and resource usage. We present an approach to SFC provisioning, consisting of three elements. The first element is a fast and scalable round-robin heuristic. The second element is a Mixed Integer Programming (MIP) based approach. The third element is a queueing-theoretic model to estimate the average latency associated with any SFC provisioning solution. Our SFC provisioning system, called Stringer, allows providers to balance the conflicting goals of minimizing infrastructure resources and end-to-end latency for meeting their respective SLAs. # nodes expected latency max utilizaton # nodes Present results to user Input data Figure 1: Flowchart of System: The input data, consisting of the service chains, their requirements and the data center network topology, are given to the heuristic. The basic solution from the heuristic provides an initial solution for the mixed integer program. Expected latency is then computed for each of the solutions from the optimizer. The user can then choose from a menu of solutions differing in expected latency and number of servers used. the specified order. For example, a service chain may require packets to follow the VNF sequence: load balancer, network address translator, and firewall. In the SFC provisioning problem, one must place (possibly multiple) instances of each VNF on servers, and choose the route(s) for each service chain, in such a way that the network can accommodate the traffic for as many service chains according to their priorities. Service chains may share VNF instances. Moreover, the traffic for a given service chain may be split among multiple paths in the network when multiple instances of a specific VNF are used.Our work differs from prior work in VNF placement in several important ways. One key difference is in the placement objective. Operators have multiple competing goals to consider when placing VNFs. A service provider may want to use as few servers as possible in order to minimize operating costs and leave open servers for future needs ([1, 2]). At the same time, the operator must ensure low end-to-end network latency for his customers. These objectives are in direct conflict. While some prior work proposes multiple alternative objectives ([3, 4, 5]), ours is the first, to our knowledge, that provides a flexible way to trade-off these competing goals in SFC provisioning. Moreover, unlike [5], our optimization model employs only linear constraints to model maximum utilization.Another important difference is...

Impact of Processing-Resource Sharing on the Placement of Chained Virtual Network Functions

“…Existing NFV node models for the estimation of expected introduced latency do not consider any processing-resource sharing aspect. The most common adopted model, which we call state-of-the-art (SOTA), merely considers a non-linear increase in the node latency as a function of the node utilization [4], neglecting any upscaling or context switching costs. In this subsection, we compare SFC embedding results obtained with our proposed processing-resource sharing node model, which relies on the costs introduced in Section III-B, with results obtained by adopting the SOTA model proposed in [4]: to do so, we have modified HCA to embed SFCs according to such simplified model.…”

“…The most common adopted model, which we call state-of-the-art (SOTA), merely considers a non-linear increase in the node latency as a function of the node utilization [4], neglecting any upscaling or context switching costs. In this subsection, we compare SFC embedding results obtained with our proposed processing-resource sharing node model, which relies on the costs introduced in Section III-B, with results obtained by adopting the SOTA model proposed in [4]: to do so, we have modified HCA to embed SFCs according to such simplified model. Specifically, considering P v = f ∈F c f,v γv as NFV node utilization, latency introduced by any NFV node v, for any VNF instance hosted by that node, is computed in the following way:…”

“…As per [4], we set K = 100 and L = 10. The computed node latencies are in the order of milliseconds, as the ones obtained with our model when ω = 0.4 ms and κ = 1.75 ms, which are the values adopted in this subsection.…”

“…Compared to existing models where resource-sharing penalties are not considered (e.g. [4]), our model enables a more accurate distribution and scaling of VNFs, preventing excessive VNF consolidation when it might jeopardize SFC performance. To the best of our knowledge, this paper is the first study evaluating the impact of such processing-resource sharing costs on service function chaining in an NFV scenario.…”

Network Slicing for Service-Oriented Networks Under Resource Constraints

“…[10]- [13] is that the instantiation of a service function for a traffic flow can be split over multiple NFV-enabled nodes. The service splitting assumption significantly simplifies the optimization problem since no binary variable is needed in the problem formulation.…”