An online distributed approach to Network Function Placement in NFV-enabled networks

Abstract

Network Function Placement (NFP) involves placing virtual network functions (VNFs) on the nodes of a network such that the data that flow through the network are processed by a chain of service functions along their path from source to destination. There are three aspects to this problem: (i) routing the flows efficiently through the network, (ii) placement of the VNFs on the nodes and (iii) steering each flow through a chain of VNFs, known as the service function chain (SFC). Routing must attempt to find “optimal” paths through the network (for e.g., shortest paths), possibly subject to constraints such as path latency and link bandwidth. The VNFs consume resources on the nodes where they are placed and are constrained by the capacity of the nodes. Steering must ensure that each flow has along its path a sequence of VNFs, likely in a certain order. One way to specify this problem is to define a multi-commodity flow problem with additional constraints based on the steering and placement requirements. Simultaneously solving all three aspects of this problem, trying to optimize various parameters and within the various constraints, is a hard problem, with even a simplified version shown to be NP-complete in this paper. Attempting to optimally solve this problem in real time while flows are getting provisioned and de-provisioned in parallel is an intractable problem, especially in large networks. Hence various types of heuristics have been used to solve this problem. In this paper we introduce a distributed, online solution that employs a message-passing protocol for nodes to negotiate the placement of the VNFs, with the minimization of the number of VNF instances being the primary objective. We compare the performance of the solution to that of the theoretically optimal solution and other proposed heuristics on both the Fat-tree topology and the BCube topology. The results show that this solution performs better than other heuristics. The average ratio of the result of the proposed solution to that of the optimal solution, taken as the approximation ratio, is found to be 1.5 for the tested scenarios.

Appendix I. NP-completeness proof of NFP problem

In this section, a simplified form of the NFP problem that involves a single network function is taken into consideration. This simplified NFP problem is termed as the Single-NFP (S-NFP) problem. It is shown that the S-NFP problem is NP-complete.

1.1 Appendix I.1 The Single Network Function Placement (S-NFP) problem

Given a network, a set of paths (representing flows through the network), a network function and a bound on the number of instances of the network function, the S-NFP problem involves finding whether the instances can be placed in the nodes of the network such that each path has at least one node (vertex) with the function placed on it. This is more formally defined here.

Let ${{\mathcal {G}}} = ({\mathcal {V}}, {\mathcal {E}})$ be an undirected graph representing a network and ${\mathcal {P}} = \{p_{1}, p_{2},\dots ,\, p_{m}\}$ be a set of m paths through the network. Each path $p_{j}$ connects two vertices $s_{j}$ and $t_{j}$, where $s_{j}, t_{j} \in {\mathcal {V}}$. The S-NFP problem is as follows: given a constant c, is there an ${\mathcal {R}} \subseteq {\mathcal {V}}$ s.t. $\mid {\mathcal {R}}\mid \;\le c$ and $\forall p_{j} \in {\mathcal {P}}, \exists {r \in {\mathcal {R}}}$ s.t. r is a vertex on path $p_{j}$?

1.2 Appendix I.2 Reduction from Set Cover problem to S-NFP

The Set Cover problem can be reduced to the S-NFP problem, proving the NP-Completeness of the S-NFP problem. The Set Cover problem is as follows. Given ${\mathcal {U}} = \{1, 2,\dots , m\}$ and a family of l subsets of ${\mathcal {U}}$, ${\mathcal {F}} = \{A_{1}, A_{2}, \dots , A_{l}\}$ and a constant c, is there a set ${\mathcal {C}} \subseteq {\mathcal {F}}$ s.t. $\mid {\mathcal {C}}\mid \le c$ and $\forall j \in {\mathcal {U}}, \exists A \in {\mathcal {C}}$ s.t. $j \in A$.

Given an instance of the Set Cover problem, a corresponding instance of the S-NFP problem can be constructed as follows.

Let ${\mathcal {U}}, {\mathcal {F}}$ and c of a Set Cover problem instance be given. A corresponding graph ${\mathcal {G}}$ and a set of m paths ${\mathcal {P}}$ through the graph can be defined based on this input. Let $s_{j}$ and $t_{j}$ be a pair of nodes in the graph corresponding to each element $j \in {\mathcal {U}}$. $s_{j}$ and $t_{j}$ ($1 \le j \le m$) are the two end nodes of path $p_{j}$ through the graph. Therefore, there are m such pairs and m corresponding paths through the graph. Additional nodes in the graph are now defined based on ${\mathcal {F}}$. For each set $A_{i} \in {\mathcal {F}}$, define a node $n_{i}$ in the graph. Also, for each $s_j-t_j$ pair, define l nodes $n^j_i$, where $1 \le i \le l$.

Each path $p_j$ in ${\mathcal {G}}$ has its end points at $s_j$ and $t_j$ and has l intermediate nodes. Intermediate node i of the path is either $n_{i}$ if $j \in A_i$, or $n^j_i$ otherwise. Essentially, all paths corresponding to elements of $A_i$ converge at $n_i$ as their $i^{th}$ intermediate node. An example of a Set Cover problem and its corresponding instance of the S-NFP problem are illustrated in figure 22.

Figure 22

Illustration of the reduction of Set Cover problem to S-NFP problem.

The NP-completeness of the S-NFP problem can be proved by proving the following lemma, thereby proving the equivalence of an instance of the Set Cover problem with the corresponding instance of the S-NFP problem.

Lemma

A solution to a feasible instance of the Set Cover problem with j elements in ${\mathcal {U}}$ and c as the bound on the cover set exists iff a solution exists to the corresponding S-NFP problem with j paths and c as the constraint on the number of network function instances.

Proof

Consider an instance of the Set Cover problem and its corresponding instance of S-NFP problem constructed as described above. Let a solution exist for the S-NFP problem, with ${\mathcal {R}}$ being the solution. The solution for the Set Cover problem instance can be constructed as follows. For each $r \in {\mathcal {R}}$, if r is node $n_i$, then include $A_i$ in the solution set ${\mathcal {C}}$. If the r is a node of the form $n^j_i$, then include any set $A_k$ s.t $j \in A_k$ (it must exist if there is a feasible solution). Since ${\mathcal {R}}$ has at least one node on each of the m paths, it follows that ${\mathcal {C}}$ has at least one set which contains each one of the m elements in ${\mathcal {U}}$. Also, $\mid {\mathcal {C}}\mid = \mid {\mathcal {R}}\mid$.

Conversely, let a solution set ${\mathcal {C}}$ exist for the Set Cover problem, solution set ${\mathcal {R}}$ for the S-NFP problem can be constructed by including in ${\mathcal {R}}$ vertex $n_i$ which corresponds to each $A_i \in {\mathcal {C}}$. Thus, above lemma is proved. It follows that S-NFP problem, and therefore the more general NFP problem, are NP-Complete. $\square$