An Opportunistic Resource Sharing and
Topology-Aware Mapping Framework for
Virtual Networks
Sheng Zhanga, Zhuzhong Qiana, Jie Wub, and Sanglu Lua
aNanjing University
bTemple University
INFOCOM 2012
Orlando, FL
March 25 – 30, 2012
1
Network Virtualization
⃝ Infrastructure provider (InP): physical/substrate network (SN)
⃝ Service provider (SP) purchases slices of resource (e.g., CPU, bandwidth,
memory) from the InP and then creates a customized virtual network
(VN) to offer value-added service (e.g., content distribution, VoIP) to end
users
2
Virtual Network Mapping
⃝ VNM is to embed multiple VN requests with resource
constraints into a substrate network
 Different virtual nodes -> different substrate nodes
 VN requests arrive over time: first come, first serve
⃝ The objective is to maximize the revenue of InP, that is,
maximize the utilization ratio of physical resources
VN request 1
VN request 2
Virtual Network Mapping
Given a VN request and a substrate nerwork, the problem of
determining whether the request can be embeded without
any constraints violation is NP-hard
[Andersen 2002]
4
Related Work
⃝ Simulated Annealing: [Ricci et al. 2003][Zhang et al. 2011]
⃝ Load Balancing: [Zhu & Ammar 2006]
 Unlimited resources
⃝ Path Splitting: [Yu et al. 2008]
 Multi-commodity flow problem
⃝ Location Constraints: [Chowdhury et al. 2009]
 Integer Linear programming + determinstic/randomized rounding
⃝ Inter-domain mapping: [Chowdhury et al. 2010]
5
Motivation
⃝ It is difficult to predict the workload precisely
 SP potentially target users all over the world
⃝ SPs often over-purchase physical resources
 To cope with a peak workload on demand
unefficient
resource
utilization
6
The ORSTA framework
1: Topology-aware node ranking (MCRank)
2: Macro level mapping
- Greedy node-to-node mapping
- maximum first
- Link-to-link mapping
- shortest path
3: Micro level assignment: for each substrate node and link,
- Local time slot assignment
7
Step 1: Topology-Aware Node Ranking
-Motivational Example
12 CPU, 8 Bandwidth
VN request 1
12 CPU, 2 Bandwidth
8
Topology-Aware Node Ranking
-Basic Idea
PageRank:
The importance of a web page is determined not only by its own
contents but also its neighbors’
Our observation:
The importance of a substrate node is determined not only by its
own resource but also its neighbors’
9
Topology-Aware Node Ranking
-Details
⃝ A node has a higher rank if it has more highly-ranked neighbors
⃝ The more neighbors one node has, the less its influence on their
rankings
Iterative effect
Actually, MCRank is the stationary distribution of a Markov chain
We prove the existence of MCRank, and also give an algorithm
for calculating it. Please refer to paper for details.
10
Step 2: Macro Level Mapping
⃝ Phase 1: node-to-node
 Sort VN nodes according to their CPU requirements
 Sort SN nodes according to their MCRank
 Maximum first matching
⃝ Phase 2: link-to-link
 shortest path
• y-z: G-H-D ?
G-F-E-D ?
 k-shortest path
• multiple edges
VN request 1
11
Step 3: Micro Level Time Slot Assignment
- Capture the fluctuation of workload
⃝ Workload model
 Basic part: always exists, its percentage is bwl
 Variable part: each unit occurs with a probability, pwl, in each time slot
⃝ CPU busy time and network flow: expressed in time slots
 proportional to the workload
Examples:
Node “x”: basic 6 + variable 6
The possible units needed: 6,7,8,…,12
bwl=0.5
pwl=0.2
12
Step 3: Micro Level Time Slot Assignment
⃝ Only focus on a substrate link
 Results can be applied to substrate nodes without any major changes
⃝ Only focus on variable sub-traffic in a substrate link
 For basic sub-traffic, we have no choice but to allocate the required number of time
slots
⃝ For variable sub-traffic
 SHARE !
13
Step 3: Micro Level Time Slot Assignment
- Tradeoff
⃝ When more than one variable sub-traffic occurs at the same time
slot, a collision happens.
⃝ To save time slots for upcoming requets
 A slot is shared among, the more virtual links the better
⃝ To guarantee performance
 A slot is shared among, the less virtual links the better
A tradeoff!
14
Step 3: Micro level Time Slot Assignment
- Breaking the tradeoff
Given multiple variable sub-traffic and a collision threshold,
find an assignment to minimize the slots used
Bin Packing
First-fit
How to accelerate the calculation of collision probability? See paper.
15
Simulation Setup
⃝ Performance metrics
 Acceptance ratio: the higher, the better
 Node/link utilization: the higher, the better
⃝ Algorithms in comparison
 ORSTA: our entire framework
 TA: only considers topology-awareness
 ORS: only considers opportunistic resource sharing
 Greedy: traditional greedy node and link mapping
16
Results: Comparison of algorithms
17
Results: Comparison of algorithms
18
Results: Comparison of algorithms
19
Results: Impacts of parameters
20
Conclusions
⃝ We re-examined the virtual network embedding problem from
two novel aspects
 Topology-awareness
 Opportunistic resource sharing
⃝ We proposed a mapping framework, ORSTA, which contains
three main components
 Topology-aware node ranking
 Macro level mapping
 Micro level time slot assignment
21
Thanks for your attention!
Q&A
http://cs.nju.edu.cn/dislab/
The Internet is a great success!
⃝Information exchange
⃝Applications support
⃝Critical infrastructure
Like many successful technologies
the Internet is suffering the adverse effects of inertia
Internet Ossification
⃝ Multiple network domains with conflicting interests
 multilateral relationship? Difficult!
 Deploy changes/updates? Global agreement!
⃝ The ever-expanding scope and scale of the Internet’s use
 security, routing stability, etc.
Flexibility + Diversity
24
Simulation Setup
Similar settings to several existing studies
⃝Substrate network
 Topology: ANSNET/Arpanet
 CPU & Bandwidth: [50,100], uniform
 Collision threshold: 0.1
⃝Virtual network
 # of nodes: [2,10], uniform
 Each pair of nodes connects with probability 0.5
 Lifetime: 10 minutes, exponential
 Arrivals: Possion process (0.2 minutes)
25
Motivation 1: example
1$ for one unit per hour
InP gets: 8$
SP1 or SP2 pays: 4$
No Free Lunch!
Collision may
happen. (0.028 here)
InP gets: (3+0.1)*3=9.3$
SP1 or SP2 or SP3 pays: 3.1$
Assumption: 4 units demand= 3 units (always needed) + 1 unit (needed with probability 0.1)
0.1$ for the shared unit per hour
26
Residual Resource Estimation
The residual room in a time slot is defined as:
the maximal probability of a variable sub-traffic that
this slot can still accommodate.
27
The ORSTA Framework
28
Topology-Aware Node Ranking
⃝ PageRank’s core idea
 A page has a higher rank if it is pointed to by more highly-ranked pages
 The more pages one page points to, the less its influence on their ranking is
⃝ MCRank
⃝ We prove that the Markov chain determined by P has a
stationary distribution, i.e., MCRank.
29