A four-stage adaptive scheduling scheme for service function chain in NFV

Abstract

Network Function Virtualization (NFV) enables the flexible software implementation of Network Functions (NFs) which is called Virtualized Network Function (VNF) and placed along the routing path of the network flow. A sequence of VNFs constitutes a Service Function Chain (SFC) to satisfy the processing requirements of flows. Since the SFC scheduling depends on the current network state and the dynamics of flows, it brings a great challenge to make an optimal SFC scheduling decision efficiently. In this paper, we present a Four-stage Adaptive Scheduling Scheme (FSASM) to make a trade-off between different scheduling goals and effects on network performance and management overhead. We design the specific mechanism for each stage when the network is in different workloads. Then, we prove the NP hardness of the optimization models in FSASM and propose a Minimum wEight Path Selection Algorithm (MEPS) with polynomial time complexity to realize a practical SFC scheduling. Moreover, we perform comprehensive experiments under different real-world topologies and network states. The results demonstrate that FSASM can achieve high network throughput and resource utilization as well as decrease the scaling frequency in highly dynamic network scenarios.

Introduction

Network Function Virtualization (NFV) seeks to implement Network Functions (NFs), e.g., network-level and application-level firewalls, Network Address Translation (NAT), proxies and Deep Packet Inspection (DPI), in software instead of special-purpose hardware appliances [1], [2], [3], which brings many benefits of high controllability, low cost and elastic scaling [4]. These software-based NFs, named Virtualized Network Functions (VNFs), can be deployed as Virtual Machines (VMs) or containers on commodity servers. When a network flow requires to be processed by several VNFs before it is forwarded to its destination [5], [6], the sequence of these VNFs constitutes a Service Function Chain (SFC) that describes the concrete processing requirement. The inappropriate SFC placement leads to the unnecessary waste of network resources, while the efficient one can take advantage of all available capacities. Hence, the transaction performance of the network depends on the scheduling results of SFCs.

Since the emergence of Software-Defined Networking (SDN) makes the management of SFCs easier through a global steering of routing paths among multiple VNFs at arbitrary locations in the network [7], [8], [9], the massive plausible scheduling candidates that satisfy the demands of all flows confuse the controller for selecting the optimal result. Furthermore, the variation of the flow rate and the dynamic of the network state sharply increase the difficulty of the SFC placement. Over the past years, a variety of researchers have made great efforts to address the SFC scheduling problem [10], [11], [12], [13], [14], [15], [16]. However, there still exist two problems. One is that these works usually concentrate on a single goal regardless of the mutable and unpredictable network situation [17], [18], which results in the improper decision and the temporary mismatching between the flow requirement and the capacity of the corresponding SFC. The other is that most studies ignore the considerable cost of elastic scaling in SFC, while it is hard to perform scaling in practice because launching a new VNF instance consumes a great deal of bandwidth and computing resources that are induced by the transfer of VM images, system booting, device attachment and state updating [19], [20], [21]. Moreover, the frequent scaling brings about the high management overhead and the long pause time. Therefore, it needs a novel scheduling scheme to efficiently overcome the performance degradation in highly dynamic network scenarios while ensuring the low management cost.

Revisiting previous SFC scheduling works, there still exists an inherent major drawback that hinders the improvement of the SFC management. As a VNF instance is overloaded, existing schemes directly trigger a scaling-out action to launch a new VNF instance [21], but the available capacity of the shared resource in the server where the VNF instance locates is not fully utilized [22], [23]. Fig. 1 illustrates the typical transmission-mismatch problem which exposes the effect of the resource utilization on the performance of the SFC management. The transmission-mismatch problem is a phenomenon that the arrival rate of data exceeds the maximum departure rate in a VNF instance, which may result from the limited processing ability of the VNF instance or the bandwidth constraint. In the multi-tenancy scenario, Fig. 1 describes a phenomenon that the SFC flows f₁ and f₂ pass through the paths {sw₁, FW, sw3, DPI, sw₄} and {sw₁, sw₂, sw₃, DPI, sw₄} respectively, while the common path, i.e., {sw3, DPI, sw₄}, becomes an inevitable bottleneck. The VNF instances along the common path may overload when the traffic burst happens on either SFC or a new SFC flow arrives. For example, either f₂ getting large due to the generation of the more SFC data or a new SFC flow f₃ being deployed on the path of f₂ will both exhaust the capacities of {sw3, DPI, sw₄}. Especially, sw3 is the critical node that suffers from the huge pressure of processing for ensuring the performance of all flows. Once the common resource is overloaded, directly dropping extra packets severely hurts the performance of SFC flows and causes the waste of the work done by previous VNF instances, but the simple scaling results in the high cost and is not necessary in the scenario of the transient overload due to the temporal traffic burst. Thus, an excellent SFC management scheme should take full advantage of the available resource in common instances to absorb the additional traffic for improving the handling efficiency and decreasing the frequency of scaling, which can not only alleviate the handling bottleneck and ensure the performance of SFC flows but also reduce management overhead.

Since assigning paths for SFC flows is the most important thing in the SFC management, we concentrate on the scheduling of SFC flows and the scaling of VNF instances. In this paper, we present FSASM, a Four-stage Adaptive Scheduling Scheme, to make a trade-off between different scheduling goals and the effects on network performance and management overhead. FSASM makes great efforts to efficiently deploy SFCs in the current network situation and perform the least but inevitable scaling events. FSASM classifies the SFC scheduling process into four stages based on different network workloads and resource utilization. In the first stage, we design the Lowest lAtency Mechanism (LAM) to take both the processing rate and the resource utilization into consideration, because the relatively small workload gives a chance to apportion the handling cost and increase the resource usage. In the second stage, with the growth of SFC demands, we propose the Dynamic rEsource Reservation Mechanism (DERM) to reduce unnecessary scaling and management overhead. In the third stage, as the VNF overload cannot be eliminated by any scheduling mechanism, we develop a novel adaptive caching mechanism to solve the transmission-mismatch problem by leveraging the shared resource in the server. In the fourth stage, once all previous mechanisms fail, FSASM falls back to directly scale the bottleneck VNF instances. While FSASM adopts SDN to achieve the SFC management, we utilize our proposed scheme, SARD [24], to reduce the number of forwarding rules and realize the fast flow-level consistent update. Furthermore, we model the SFC scheduling problems in LAM and DERM by Integer Linear Programming (ILP) as well as prove their NP hardness. And we devise the Minimum wEight Path Selection Algorithm (MEPS) with polynomial time complexity to solve the NP-hard problems in FSASM. We perform comprehensive experiments in various network scenarios. The results demonstrate that FSASM can obviously improve network throughput and decrease the scaling frequency, especially under highly dynamic network situations.

Our main contributions are summarized as follows:

• We present a Four-stage Adaptive Scheduling Scheme that meets the requirements of highly dynamic flows and achieves high network performance while reducing unnecessary scaling events and the management cost by leveraging different mechanisms in each stage.

• We formulate the SFC scheduling problems in FSASM by using Integer Linear Programming and devise an efficient algorithm to solve the corresponding problems which are proven as NP-hard.

• We conduct comprehensive experiments in various network scenarios to demonstrate the efficiency of our proposed designs.

The rest of this paper is organized as follows. In Section 2, we discuss related works. In Section 3, we detail the models of network and SFC. We elaborate on FSASM in Section 4. In Section 5, we prove the NP hardness of the scheduling problems in FSAM and present MEPS to solve these problems. In Section 6, we evaluate the performance of our designs. Section 7 concludes our work and provides future directions.