This document discusses mobile cloud computing and data center network architectures. It begins with an introduction to mobile cloud computing, including definitions and key elements. It describes how mobile devices can offload computation and data to augment their capabilities. It then covers mobile networks and technologies for cloud access. The document outlines approaches to distributed application processing and partitioning. Finally, it defines data center network architectures and classification schemes before analyzing examples like Fat-Tree, BCube and DCell. Evaluation criteria for energy consumption and communication latency in data centers are also presented.

1. University of Luxembourg
Distributed systems and middleware
Claudio Fiandrino
claudio.fiandrino@uni.lu
November 17, 2014

2. Agenda
Mobile Cloud Computing
Data Center Network Architectures
Claudio Fiandrino | Agenda - 1 of 37

3. Scenario
Mobile Operator Network
CloudInternetP-GWS-GW
MME
HSS PCRFeNodeB
IP Network
WiFi AP
WiFi APUEs
Claudio Fiandrino | Agenda - 2 of 37

4. Outline
Mobile Cloud Computing
Introduction
Mobile Devices
Mobile Networks
Distributed Application Processing
Data Center Network Architectures
Claudio Fiandrino | Mobile Cloud Computing - Introduction 2 of 37

5. Mobile Cloud Computing
Mobile Cloud Computing vs. Mobile Computing
In Mobile Cloud Computing the devices run cloud-based apps
In Mobile Computing the devices run native apps
In Mobile Cloud Computing users have to access remotely stored
applications and their associated data over the Internet
Claudio Fiandrino | Mobile Cloud Computing - Introduction 3 of 37

6. Mobile Cloud Computing
Mobile Cloud Computing vs. Mobile Computing
In Mobile Cloud Computing the devices run cloud-based apps
In Mobile Computing the devices run native apps
In Mobile Cloud Computing users have to access remotely stored
applications and their associated data over the Internet
Mobile Cloud Computing: Common Understanding
Access to cloud computing services through mobile devices
Claudio Fiandrino | Mobile Cloud Computing - Introduction 3 of 37

7. Mobile Cloud Computing: Definition
Definition
An integration of cloud computing technology with mobile devices to
make the mobile devices resource-full in terms of computational power,
memory, storage, energy, and context awareness.
[1] Khan, A.R.; Othman, M.; Madani, S.A.; Khan, S.U., “A Survey of Mobile Cloud Computing Application
Models," IEEE Communications Surveys & Tutorials, vol.16, no.1, pp.393,413, First Quarter 2014
doi: 10.1109/SURV.2013.062613.00160
Claudio Fiandrino | Mobile Cloud Computing - Introduction 4 of 37

8. Mobile Cloud Computing: Definition (II)
Definition (II)
Mobile cloud computing is a model for transparent elastic augmentation
of mobile device capabilities via ubiquitous wireless access to cloud
storage and computing resource, with context-aware dynamic adjusting
of offloading in respect to change in operating conditions, while
preserving available sensing and interactivity capabilities of mobile
devices.
[1] Dejan Kovachev, Yiwei Cao, Ralf Klamma: Mobile Cloud Computing: A Comparison of Application
Models. CoRR, abs/1107.4940, 2011.
(http://arxiv.org/abs/1107.4940v1)
Claudio Fiandrino | Mobile Cloud Computing - Introduction 5 of 37

9. Keywords
Mobile devices
Cloud Computing
Augmentation
Preserving capabilities
Resource-full
Claudio Fiandrino | Mobile Cloud Computing - Introduction 6 of 37

10. Key Elements
Mobile
Cloud
Computing
Mobile
Devices
Network
Cloud
Applications
Claudio Fiandrino | Mobile Cloud Computing - Introduction 7 of 37

11. Outline
Mobile Cloud Computing
Introduction
Mobile Devices
Mobile Networks
Distributed Application Processing
Data Center Network Architectures
Claudio Fiandrino | Mobile Cloud Computing - Mobile Devices 7 of 37

12. Type of Devices
Not only smartphones and tablets
Claudio Fiandrino | Mobile Cloud Computing - Mobile Devices 8 of 37

13. Issues of mobile devices
Resource scarceness
Battery constrained
Low connectivity
Claudio Fiandrino | Mobile Cloud Computing - Mobile Devices 9 of 37

14. Solution
Offloading to the Cloud
Computation
Data
Outcome: Application partitioning
Cope with mobile device issues
Gather more data
Use idle processing power
Claudio Fiandrino | Mobile Cloud Computing - Mobile Devices 10 of 37

15. Outline
Mobile Cloud Computing
Introduction
Mobile Devices
Mobile Networks
Distributed Application Processing
Data Center Network Architectures
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 10 of 37

16. Technologies
Cellular (3G/4G)
WiFi
Bluetooth
Comparison
TECHNOLOGY MAX DATA RATE ENERGY RANGE SPECTRUM
Cellular Up to 300 Mb/s1 High More than 10 km (1 km)2 Licensed
WiFi (802.11n) Up to 100 Mb/s Medium Up to 250 m (120 m) Unlicensed
Bluetooth Up to 3 Mb/s Low Up to 100 m (20-30 m) Unlicensed
1. Considering 4x4 MIMO, 20 MHz channel
2. Macro cell radius and typical urban cell radius
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 11 of 37

17. Cellular Coverage
http://opensignal.com/
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 12 of 37

18. Health State of the Global Mobile Traffic
4.4 billion people will use mobile cloud applications by 2017
Mobile data traffic will reach 15 EB (1018
) per month in 2018
22 million wearable devices generated 1.7 PB (1015
) per month traffic
in 2013
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 13 of 37

19. What about Energy?
The importance of the technologies
“Our calculations show that, in 2015, the wireless networks we use to
access cloud services will command around 90% of the energy needed
to power the entire wireless cloud services ecosystem. By comparison,
data centres will account for only 9% or less. Industry needs to focus on
the real issues with wireless network technologies if it wants to solve this
problem.”
CEET (Centre for Energy-Efficient Telecommunications), Massive energy cost hidden in wireless cloud
boom, 2013
(http://www.ceet.unimelb.edu.au/news/media/2013-04-09.html)
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 14 of 37

20. What about Energy?
Cloud Wireless Access Burns More Energy Than Data Centres – Report
(http://www.techweekeurope.co.uk/workspace/cloud-wireless-energy-use-data-centre-113227)
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 14 of 37

21. What about Energy? (II)
Interesting Statistics
One Google search is equivalent to about 0.2 grams of CO2
(http://googleblog.blogspot.com/2009/01/powering-google-search.html)
Raffi (Twitter API team) says: “One tweet consumes around 100 J”
http://goo.gl/uMnV0f
http://www.slideshare.net/raffikrikorian/energy-tweet
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 15 of 37

22. Diagnosis
Networking can not be neglected
Communications may impact dramatically on performance
Carefully select proper technologies
Claudio Fiandrino | Mobile Cloud Computing - Mobile Networks 16 of 37

23. Outline
Mobile Cloud Computing
Introduction
Mobile Devices
Mobile Networks
Distributed Application Processing
Data Center Network Architectures
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 16 of 37

24. Application Partitioning: Objectives
Enhance mobile devices’ performance
Energy saving
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 17 of 37

25. High level list of applications
Commerce
Healthcare
Gaming
Searching (Keyword, Voice, Image, Location, Tag)
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 18 of 37

26. Application Partitioning Classification
Static Partitioning
A set of task is always run remotely
Dynamic Partitioning
Tasks are run where it is most convenient
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 19 of 37

27. Where?
Mobile Operator Network
CloudInternetP-GWS-GW
MME
HSS PCRFeNodeB
IP Network
WiFi AP
WiFi APUEs
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 20 of 37

28. Where?
Mobile Operator Network
CloudInternetP-GWS-GW
MME
HSS PCRFeNodeB
IP Network
WiFi AP
WiFi APUEs
Remote Clouds
Local Clouds
Other mobile devices
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 20 of 37

29. A practical example: Face Recognition
Flow
Detection on picture
Determine match on database
Partitioning
1. Local detection/remote recognition
2. Remote detection/remote recognition
Beware: what is remote?
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 21 of 37

30. List of Approaches
CloudLets
CloneCloud
MAUI
MobiCloud
Odessa
Cuckoo
Hyrax
µCloud
ThinkAir
eXCloud
[1] Khan, A.R.; Othman, M.; Madani, S.A.; Khan, S.U., “A Survey of Mobile Cloud Computing Application
Models," IEEE Communications Surveys & Tutorials, vol.16, no.1, pp.393,413, First Quarter 2014
doi: 10.1109/SURV.2013.062613.00160
[2] Shiraz, M.; Gani, A.; Khokhar, R.H.; Buyya, R., “A Review on Distributed Application Processing
Frameworks in Smart Mobile Devices for Mobile Cloud Computing,” IEEE Communications Surveys &
Tutorials, vol.15, no.3, pp.1294,1313, Third Quarter 2013
doi: 10.1109/SURV.2012.111412.00045
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 22 of 37

31. Take-home messages
Mobile Cloud Computing
Augmenting devices capabilities
Heterogeneity (applications/technologies/user behaviour)
Application Partitioning
To keep in mind
Not simply “Access to cloud computing services through mobile devices.”
Claudio Fiandrino | Mobile Cloud Computing - Distributed Application Processing 23 of 37

32. Outline
Mobile Cloud Computing
Data Center Network Architectures
Definition and Classification
Architectures Analysis
Energy Consumption and Communication Latency Evaluation
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 23 of 37

33. Data Center Definitions
Data Center Infrastructure
The data center is home to the computational power, storage,
management and dissemination of data and information necessary to a
particular body of knowledge or pertaining to a particular business.
Data Center Network Architecture
The data center network architecture is the set of network nodes and
links that characterize the interconnectivity among the computing
servers and to the external world.
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 24 of 37

34. Design Criteria
Scalability
Flexibility
Resiliency
Maintenance
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 25 of 37

35. A brief list of Architectures
Fat-Tree
“Al-Fares”
Portland
Hedera
Clos (VL2)
Hypercube
BCube
DCell
DPillar
FiConn
FlatNet
Helios
C-Through
Petabit
GreenCloud (Not the Simulator)
[1] Ali Hammadi, Lotfi Mhamdi, A survey on architectures and energy efficiency in Data Center
Networks, Computer Communications, Volume 40, 1 March 2014, Pages 1-21, ISSN 0140-3664,
http://dx.doi.org/10.1016/j.comcom.2013.11.005.
(http://www.sciencedirect.com/science/article/pii/S0140366413002727)
[2] Bari, M.F.; Boutaba, R.; Esteves, R.; Granville, L.Z.; Podlesny, M.; Rabbani, M.G.; Qi Zhang; Zhani,
M.F., “Data Center Network Virtualization: A Survey," IEEE Communications Surveys & Tutorials, vol.15,
no.2, pp.909,928, Second Quarter 2013 doi: 10.1109/SURV.2012.090512.00043
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 26 of 37

36. Classifications
Electronic/Optical
Which technology is used in the forwarding?
Switch/Server Centric
Who performs the forwarding?
Ali Hammadi, Lotfi Mhamdi, A survey on architectures and energy efficiency in Data Center Net-
works, Computer Communications, Volume 40, 1 March 2014, Pages 1-21, ISSN 0140-3664,
http://dx.doi.org/10.1016/j.comcom.2013.11.005.
(http://www.sciencedirect.com/science/article/pii/S0140366413002727)
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 27 of 37

37. Electronic vs Optical
Fully electronic
Fully optical
Hybrid
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 28 of 37

38. Swich Centric vs Server Centric
Switch Centric: switches are the key components in forwarding
Switches are complex
Servers no forwarding functionalities
Bandwidth oversubscription
Server Centric: servers are the key components in forwarding
Switches are very simple
Servers with forwarding functionalities
No bandwidth oversubscription
Trade off
Servers devoted to computational purposes only
Exploiting full link capacity
Claudio Fiandrino | Data Center Network Architectures - Definition and Classification 29 of 37

39. Outline
Mobile Cloud Computing
Data Center Network Architectures
Definition and Classification
Architectures Analysis
Energy Consumption and Communication Latency Evaluation
Claudio Fiandrino | Data Center Network Architectures - Architectures Analysis 29 of 37

40. Three-tier
Hierarchical structure
Bandwidth oversubscription
Computing Servers
Access Layer
Aggregation Layer
Core Layer
Gateway Router
Internet
Claudio Fiandrino | Data Center Network Architectures - Architectures Analysis 30 of 37

41. BCube
Main parameters:
n: number of ports per switch
k + 1: number of ports per server
Number of servers: nk+1
Number of switches: (k + 1) · nk
Example with n = 4 and k = 1
Internet
Computing Servers
Commodity Switches
Level 0
Commodity Switches
Level k + 1
Load Balancers
Gateway Router
Claudio Fiandrino | Data Center Network Architectures - Architectures Analysis 31 of 37

42. DCell
Main parameters:
n: number of servers per DCell0
k: number of DCell levels, server
degree of connectivity (k + 1)
Number of servers:
tj+1 = tj · (tj + 1), t0 = n
Number of switches (#DCell0):
gj = tj−1 + 1 or #servers/n
Example with n = 3 and k = 1
Internet
Computing Servers
Commodity Switches
Load Balancers
Gateway Router
Claudio Fiandrino | Data Center Network Architectures - Architectures Analysis 32 of 37

43. Outline
Mobile Cloud Computing
Data Center Network Architectures
Definition and Classification
Architectures Analysis
Energy Consumption and Communication Latency Evaluation
Claudio Fiandrino | Data Center Network Architectures - Energy Consumption and Communication Latency Evaluation 32 of 37

44. Evaluation Criteria: Computing Servers
Power consumption (DVFS)
P(l) = Pidle +
Ppeak − Pidle
2
· (1 + l − e−( l
τ ))
l server load
τ in the range [0.5, 0.8]
Key idea:
Load= 0
Servers
Access Layer
Aggregation Layer
Core Layer
→ →
Increasing Load
→ →
Load= 1
Idle link; Active link; Idle device; Active device;
Number of computing servers: 4096
Idle and peak power from Dell PowerEdge R720, Huawei Tecal
RH2288H V2 and IBM System x3500 M4
Claudio Fiandrino | Data Center Network Architectures - Energy Consumption and Communication Latency Evaluation 33 of 37

45. Evaluation Criteria: Network
Three-tier:
128 racks
16 aggregation switches
8 core switches
BCube:
n = 8, k = 3
2048 switches
DCell:
n = 8
2 < k < 3
Other parameters:
1 Gbits link (Three-Tier, BCube and
DCell)
10 Gbits link (Three-Tier)
Test packets of 40 B and 1500 B
Methodology
One way delay
Transmission and Propagation
Claudio Fiandrino | Data Center Network Architectures - Energy Consumption and Communication Latency Evaluation 34 of 37

46. Energy: Results
0 0.2 0.4 0.6 0.8 1
0
1
2
3
4
·105
Load (l)
PowerConsumption(W)
Dcell Three-tier BCube
Claudio Fiandrino | Data Center Network Architectures - Energy Consumption and Communication Latency Evaluation 35 of 37

47. Communication: Results
PERFORMANCE INDEX
ARCHITECTURES
Three-tier BCube DCell
Latency (40 B) 1.98 µs 3.93 µs 4.73 µs
Latency (1500 B) 28.34 µs 73.72 µs 93.92 µs
Hop distance 5.78 7.00 8.94
Server Degree Connectivity 1 4 2.79
Beware: we counted as 2 the number of links between any pair of servers within a DCell0 and in BCube.
Claudio Fiandrino | Data Center Network Architectures - Energy Consumption and Communication Latency Evaluation 36 of 37

48. Take-home messages
Data Center Network Architectures
Different Architectures have:
Different impact on energy costs
Different impact on communication processes
Claudio Fiandrino | Data Center Network Architectures - Energy Consumption and Communication Latency Evaluation 37 of 37

49. Thank You!Thank You!Thank You!