Optimizing Flow Bandwidth Consumption with Traffic-diminishing - PowerPoint PPT Presentation

Optimizing Flow Bandwidth Consumption with Traffic-diminishing Middlebox Placement Yan Yang Chen en, Jie Wu, and Bo Ji Center for Networked Computing Temple University, USA

VNF: Evolution of Network Service l Network Function Virtualization (NFV) ¡ Virtualizing network functions into software building blocks l Virtualized Network Function (VNF) or Middlebox ¡ Software implementation of network functions ¡ Improve performance & enhance security l Examples Firewall NAT Proxy l Middlebox Deployment ¡ Deployment location selection on multiple servers

VNF Traffic Changing Effects [1] l VNFs may change flow rates in different ways ¡ Citrix CloudBridge WAN accelerator: 20% (diminishing) ¡ BCH(63,48) encoder: 130% (expanding) Checksum Data 1 0 1 1 1 0 1 0 [1] Traffic Aware Placement of Interdependent NFV Middleboxes (INFOCOM ’17)

A motivating example Initial flow rate: Traffic-diminishing ratio f 1 (4), f 2 (2), f 3 (2), f 4 (2) of VNF m: 0.5 0.5*4*2+ 0.5* 2*2+ 0.5* 2+ 0.5* 2=8 0.5*4*2+2*2+2+2=12 Total bandwidth consumption

2. Our model l Problem ¡ Deploy a single type of VNFs with traffic- diminishing effect into the network l Objective ¡ Minimize total bandwidth consumption of all flows on all links along their paths l Constraint ¡ Each flow gets processed ¡ Deploy a limited number of the single type of VNFs

3. Problem Formulation A mathematical optimization problem on minimizing total flow bandwidth consumption Single flow

4. Solution for general topologies l NP-hard l Decrement function ¡ Decrement of total bandwidth consumption compared to no VNFs l Marginal decrement ¡ Additional bandwidth decrement by deploying on 𝒯 beyond 𝒬 l Decrement function is submodular ¡ More VNFs, less bandwidth consumption ¡ Flow gets processed no later than 𝒬

4. Solution for general topologies (cont’d) l Solution ¡ General Topology Placement (GTP) l Steps ¡ Iteratively select v ∈ V with the maximum marginal decrement until all flows are fully served ! l Approximation ratio 1 − " l Time complexity ( |V|: #vertices ) ¡ O(|V| 2 log |V |)

5. Two solutions for trees Solution 1: Dynamic Programming (DP) l 𝐺(𝑤, 𝑙) ¡ Minimum total occupied bandwidth of all flows with 𝑙 deployed middleboxes in subtree 𝑈 v rooted at 𝑤 ¡ All flows get fully processed in T v l 𝑄(𝑤, 𝑙, 𝑐) ¡ Same as F(v,k) ¡ When flows with only a total bandwidth 𝑐 processed l Optimal solution l Time complexity (|V|: #node, 𝑠 !"# : largest flow rate) ¡ 𝑃(|𝑊| (log |𝑊|) ! 𝑠 "#$ )

Solution 1: Dynamic Programming (DP) Partially processed (a) Subtree fully processed (b) Processed on v Fully processed

Solution 2: Heuristic Algorithm for Trees (HAT) l Lowest Common Ancestor (LCA) ¡ LCA(v,w): lowest vertex have both v and w as descendants l Steps ¡ Deploy one VNF on each leaf vertex ¡ Delete two VNFs on v and w with minimum difference of the total bandwidth value ¡ Place one VNF on LCA(v,w) ¡ Until total number of deployed VNFs no more than k

4. Solution for trees (cont’d) l Maintenance of all difference values ¡ Min-heap ¡ Improve time efficiency l Time complexity ¡ O(|V | 2 log |V |) ¡ |V|: #vertices

7. Simulation l Comparison algorithms ¡ Random l Randomly deploy k VNFs ¡ Best-effort l Deploy on the vertex, which can reduce the total bandwidth of flows most, until k VNFs are deployed l Our proposed algorithms ¡ General topo l Alg. GTP ¡ Tree topo l Algs. GTP, DP, HAT

Settings l Topology l Middlebox traffic-diminishing ratio ¡ From 0 (e.g., spam filters) to 0.9 (e.g., traffic optimizer) with a stride of 0.1 ¡ Additional simulation on spam filter l Flow rate distribution ¡ CAIDA data center 1-hour packet trace

Simulation results of tree 10 5 10 5 l Alg. DP performs 1.6 2.5 Random Random Bandwidth consumption best for all four Best-effort Bandwidth consumption Best-effort 1.4 GTP GTP 2 variables HAT HAT DP 1.2 DP l k = 1, only one 1.5 1 feasible placement 1 0.8 plan for all methods 0.6 0.5 5 10 15 l Traffic-changing 0 0.2 0.4 0.6 0.8 k Traffic-changing ratio 10 5 10 5 ratio has the 2 3 Random Random largest impact on Bandwidth consumption Bandwidth consumption Best-effort Best-effort 1.8 2.5 GTP GTP the bandwidth HAT HAT 1.6 DP DP 2 consumption 1.4 1.5 1.2 l Random has the 1 1 biggest fluctuation 0.8 0.5 0.3 0.4 0.5 0.6 0.7 0.8 15 20 25 30 Flow density Topology size Tree Topology

Simulation results of general topology l Alg. GTP always 10 5 10 5 5.5 consumes the smallest Random Random Bandwidth consumption Best-effort Bandwidth consumption 4 Best-effort GTP bandwidth GTP 5 3.8 l Error bars become 3.6 4.5 3.4 shorter 4 3.2 l Bandwidth consumption 3 3.5 increases faster in fig. 12 14 16 18 20 22 0 0.2 0.4 0.6 0.8 k Traffic-changing ratio b when ratio ranges 10 5 10 5 8.0 8.0 from 0.4 to 0.6 Random Random Bandwidth consumption Bandwidth consumption Best-effort Best-effort GTP GTP l When flow density is 6.0 6.0 lower than 0.4 in fig. c, little difference among 4.0 4.0 three algorithms 2.0 2.0 0.3 0.4 0.5 0.6 0.7 0.8 20 30 40 50 Flow density Topology size General Topology

Simulation results (cont’d) Spam Filter (Traffic diminishing ratio: 0) l Flow density plays a more important role in affecting the total bandwidth consumption l When flow density doubles from 0.3 to 0.6, bandwidth consumption in tree increases 30.2%, while increment is only 25.6% in general topo

Conclusion and Future Work l Problem ¡ Deploy a limited number of traffic-diminishing VNFs ¡ All flows get processed l Objective ¡ Minimize total bandwidth consumption l Solutions ¡ Tree: optimal and greedy ¡ General graph: performance-guaranteed l Future Work ¡ Traffic-expanding VNFs ¡ Service chain: an ordered set of multiple VNFs

Optimizing Flow Bandwidth Consumption with Traffic-diminishing Middlebox Placement Thank you!