Project 1-Task 1 Routing Performance in Distributed SDC under - - PowerPoint PPT Presentation

Project 1-Task 1

PhD students: Ziyao Zhang and Qiaofeng Qin PI and Collaborators: Liang Ma, Konstantinos Poularakis, Kin Leung, Leandros Tassiulas, Dave Conway-Jones, Andreas Martens, Franck Le, Sastry Kompella, and Jeremy Tucker

Imperial, Yale, IBM US/UK, NRL, Dstl

Routing Performance in Distributed SDC under Synchronization Constraint

P1T1 US-UK Collaborations

P1T1 Distributed SDC

Theory

Ziyao Zhang, Liang Ma, Konstantinos Poularakis, Kin Leung, Leandros Tassiulas, Franck Le, Sastry Kompella Imperial, IBM US, Yale, NRL

Experiment

Qiaofeng Qin, Konstantinos Poularakis, Leandros Tassiulas, Sastry Kompella, Andreas Martens Yale, NRL, IBM UK

Demo

Qiaofeng Qin, Konstantinos Poularakis, Leandros Tassiulas, Kin Leung, Sastry Kompella, Andreas Martens, Dave Conway-Jones, Franck Le, Jeremy Tucker Yale, Imperial, IBM US/UK, NRL, Dstl

Modeling

Ziyao Zhang, Liang Ma, Konstantinos Poularakis, Kin Leung, Leandros Tassiulas, Franck Le, Sastry Kompella, Jeremy Tucker Imperial, IBM US, Yale, NRL, Dstl

2

Motivations

Software Defined Coalitions (SDC)

• programmable coalition

management

• easy reconfiguration
• n-demand resource allocation
• rapid response to failures in

military networks enclave enclave enclave enclave Challenges in realizing distributed SDC

• fundamental understanding of key factors affecting the
• verall performance in distributed SDC

Ø Experiments Ø Analytical Results

3

Distributed SDC

Experiments

§ Implement a real wireless SDN system by installing commercial SDN

components (ONOS controller, Open vSwitch data path) into mobile devices (Android smartphones)

§ Measure the delay of SDN control (time required to reconfigure a data plane

device by its nearest controller):

– The average delay is highly sensitive to #hops and the controller placement strategy

Controller Controller Wi-Fi Hotspot

• Android Smartphone

ONOS Open vSwitch Wi-Fi Interface

4

Large-scale Emulations

§ Emulate networks with hundreds of nodes using Mininet virtual testbed. § Two types of SDN control overheads measured:

– Controller-node traffic to collect state information and send flow setup rules. – Inter-controller traffic to synchronize the states among controllers. – The two types of overhead are significant and of the same order of magnitude

(up to a few Mbps).

– Increase almost linearly with the network size (#nodes/flows/controllers).

Controller Cluster Mininet

Edge Node 1

· · · · · · Inter-Controller Traffic Controller-Node Traffic

ONOS ONOS ONOS Edge Node 2 Edge Node 3 Edge Node 4 Edge Node 5 Edge Node N

5

Impact of Synchronization on Network Performance

§ The above experiments/emulations highlight the feasibility of

distributed SDN and the operational overheads due to synchronizations

§ challenging to achieve full status synchronizations

among controllers in a real system

• we therefore focus on analyzing and quantifying the

network performance in distributed SDN, under incomplete synchronization levels and different network structural properties 6

Problem Formulation – Network Model (two-layer model)

top-layer

• vertex: domain
• edge: two domains are connected

bottom-layer physical intra-/inter-domain connections constrained by the top- layer topology A generic model:

• input: #domains, #nodes/#gateways in each domain, top-layer

vertex degree distribution, and intra-domain node degree distribution

• no assumptions on specific random graph models

7

Problem Formulation - Performance under Dynamically Adjusted Link Preference Levels

§ Link preference level: SDN controllers assign weight (called link

preference) to links according to traffic status, security policies, and

• ther collected network information

– smaller the link weight à better for path construction – link preference is dynamically adjusted by controllers à modeled as

random variables

§ no assumptions on the pdf of these random variables

– Goal: find the path that incurs the minimum accumulated end-to-end path

cost between two arbitrary nodes in different domains (average path cost - APC)

Objective: derive APC expression under various link preference levels and inter-domain synchronization scenarios when a basic routing mechanism is used 8

Synchornization Scenarios

Highest Level Basic Sync Minimum Sync (MS) Self-domain Sync (SS) Partial Sync (PS) Complete Sync (CS)

Highest sync level: generally cannot achieve

§ Synchronization: domain A is synced with B if A knows

the minimum path cost for any two nodes in B

§ Synchronization radius: max integer τ à all domains

within τ-1 hops in the domain-wise topology are synced

• nly domain reachability

and intra-domain topology without link preference are known intra-domain link preference is known; no inter-domain sync (τ=1) sync level between SS and CS; quantified by the synchronization radius each domain is synced with all other domains

9

ü

Path construction mechanism - RCPC

a new path construction mechanism a basic and representative mechanism

RCPC (Routing Cluster based Path Construction)

• Select the shortest domain-wise path (e.g., BGP-like protocols)
• According to the given synchronization level (synchronization

radius), partition domains on this domain-wise path into routing clusters (RCs) – e.g., sync radius=2

• Construct the shortest path segment in each RC
• Concatenate all path segments into a path

RC1 RC2 RC3

10

Asymptotic Analysis

Theorem 1. Given the synchronization radius τ, the asymptotic APC (L) in the two- layer network model is Network Structural Parameters:

• m/n/𝛿 : #domains, #nodes/#gateways in each

domain

• ζi: average #vertices that are i-hop away

from a random vertex in an auxiliary graph à related to degree distributions

• Δ: average domain-wise distance wrt two

domains ~ log(m) Synchronization-related Parameters:

• τ’ = min(τ, Δ+1)
• It captures dominant parameters
• Theorem 1 applies to any sync

scenario

• When τ is smallà minimum sync,

Theorem 1 is reduced to ~ O(m*log(n)).

• When τ is large à complete sync,

Theorem 1 is reduced to ~ O(log(n*log(m))).

11

Asymptotic Analysis – cont’d

Theorem 1. Given the synchronization radius τ, the asymptotic APC (L) in the two- layer network model is Network Structural Parameters:

• m/n/𝛿 : #domains, #nodes/#gateways in each

domain

• ζi: average #vertices that are i-hop away

from a random vertex in an auxiliary graph à related to degree distributions

• Δ: average domain-wise distance wrt two

domains ~ log(m) Synchronization-related Parameters:

• τ’ = min(τ, Δ+1)
• APC reduction declines with the increase of τ

(sync level) & the increase of 𝛿 (#gateways)

• Performance and cost trade-off in distributed SDC

12

Fine-grained APC Expressions

• Fine-grained accurate APCs à for link preference levels that are dynamically

adjusted by the controllers based on up-to-date network status information

• Methodology Sketch:

Ø Given the distribution of the link preference levels, compute the intra-domain path cost distribution using mixture distribution Ø Compute the minimum path cost between an arbitrary node and the closest gateway within each routing cluster Ø Add the costs of all path segments in all traversed routing clusters

Results are widely applicable and accurate

• input: #domains, #nodes/#gateways in each domain, top-layer

vertex degree distribution, and intra-domain node degree distribution

• no assumptions on specific random graph models
• no assumptions on the distribution of link preference levels

13

Evaluations

• Data sources:

Ø Real network traces: CAIDA, Routeview, and Rocketfuel data Ø Synthetic networks: Erdos-Renyi and Barabasi-Albert network models

• Methodology:

Ø Use the network parameters extracted from the above real and synthetic data sources Ø Randomly select src/dst pairs; compare the real against the analytical APCs

• APC changes with varying parameters are closely captured by the asymptotic analysis
• APC is a decreasing convex function of the number of gateways

à diminishing benefits when the inter-domain connection is dense

• Increasing #gateways is more effective in reducing APC than increasing the

synchronization radius

• The network size n does not have a significant impact on APC

#gateways sync radius #nodes in each domain

14

Evaluations – Cont’d

• the simulation curves are closely approximated by the theoretical results for all

synchronization scenarios

• higher synchronization level is beneficial in reducing APCs

Ø SS (self-domain sync) outperforms MS (minimum sync) by ~30%; CS (complete sync) outperforms MS by ~70%

• when link preference levels are random variables

15

Relation to Other Projects

P1T1 distributed SDC foundations

Design Inspiration P2 on policy design P3 on functional entity placement P3/P4 on resource management P1T2 on SDC systems

16

External Publications

§ [1] K. Poularakis, Q. Qin, E. Nahum, M. Rio, and L. Tassiulas, "Extending SDN

Control through Programmable Mobile Devices", Workshop on Distributed Analytics InfraStructure and Algorithms for Multi-Organization Federations, proc. of IEEE Smart World Congress, 2017.

§ [2] Q. Qin, K. Poularakis, G. Iosifidis, and L. Tassiulas, "SDN Controller Placement at

the Edge: Optimizing Delay and Overheads", IEEE INFOCOM, 2018.

§ [3] K. Poularakis, Q. Qin, E. Nahum, M. Rio, and L. Tassiulas, "Flexible SDN Control

in Tactical Ad Hoc Networks", Elsevier Ad Hoc Networks Journal (to appear)

§ [4] K. Poularakis, Q. Qin, L. Ma, S. Kompella, K. K. Leung, and L. Tassiulas,

“Learning the Optimal Synchronization Rates in Distributed SDN Control Architectures” (Submitted to IEEE INFOCOM 2019, under review)

§ [5] Z. Zhang, L. Ma, K. K. Leung, F. Le, S. Kompella, and L. Tassiulas, “How

Advantageous Is It? An Analytical Study of Controller-Assisted Path Construction in Distributed SDN” (Submitted to ACM SIGMETRICS 2019, under review)

§ [6] Z. Zhang, L. Ma, K. K. Leung, and F. Le, “More Is Not Always Better: An Analytical

Study of Controller Synchronizations in Distributed SDN” (Submitted to IEEE INFOCOM 2019, under review)

17

Backup Slides

18

Universal Lower Bound

Theorem 2. Universal APC lower bound: Given the synchronization radius τ, the lower bound of APC in the two-layer network model is § 𝑀$%&'( is a logarithmic function non-increasing with τ § Theorem 2 applies to networks with any sync and link preference levels § When the number of gateways in each domain is sufficiently large

à 𝑀$%&'( is significantly simplified due to easier inter-domain routing

19