Beruflich Dokumente
Kultur Dokumente
Systems Perspective
Prof. Nalini
Venkatasubramanian
Department of Computer Science
University of California, Irvine
Outline
Users
Applications
Networks Mobile
Devices
Access Point WLAN
Laptop
Cell Phone
ad hoc
BS
network
GSM or PDA
CDMA
Technology Incentive
MOBILE HOST
BASE
STATION FIXED
BASE HOST
WIRELESS RADIO CELL
STATION
9Kbps - 14Kbps
FIXED NETWORK
BASE
STATION
Mbps to Gbps
BASE
STATION
PDA
Wireless network
characteristics
Variant Connectivity
Low bandwidth and reliability
Frequent disconnections
predictable or sudden
Asymmetric Communication
Broadcast medium
Monetarily expensive
Charges per connection or per message/packet
Location changes
location management - cost to locate is added to communication
Heterogeneity in services
bandwidth restrictions and variability
Dynamic replication of data
data and services follow users
Querying data - location-based responses
Security and authentication
te l
on a
xt
Changing global system state t e t io
C lob
(audio/video)
G xt n
Synchronization (e.g. lip sync., floor
control)
Challenges
ce t
N on
i
et te
C
x
ev te
w xt
D n
or
o
k
C
Need high degree of network Heterogeneous devices
awareness and Limited battery lifetime
customizability Size/weight limitations
congestion rates, mobility patterns Computation/Communic
etc. ation constraints
QoS driven resource provisioning
Heterogeneous networks
Mobile Multimedia Rich Media
applications on mobile devices
OBJECTIVE : Stream rich multimedia content at highest possible quality
(user experience) over wired and wireless networks
Low-power
mobile device
Video stream
Wide Area Wireless
Network Network
MEDIA SERVER Access point Video request
Challenges
Soft Real time Requirements
High demands on CPU / Network
Loss in performance directly affects user perception
Opportunities
Predictable regular behavior allows for interesting optimizations and
adaptations
Multimedia Information
Systems: Challenges
System
Generally used to
(Operating and Communication System)
express constraints on
(System QoS)
Timing, availability, (Network QoS)
(Device QoS)
reliability,security,
resource utilization MM devices Network
QoS Classes
QoS Service Classes determine
reliability of offered QoS
utilization of resources
Guaranteed Service Class
QoS guarantees are provided based on deterministic and statistical QoS
parameter values.
Predictive Service Class
QoS parameter values are estimated and based on the past behavior of the
service
Best-effort Service Class
No guarantees or only partial guarantees provided
No QoS parameters are specified or some minimal bounds are given.
Supporting continuous
media: Approaches
Admission control
Provide different service classes
Fast storage and retrieval of Multimedia content
Optimized organization (placement) of multimedia files
on disk
Special disk scheduling algorithms
Efficient Memory Management
Sufficient buffers to avoid jitter
Intelligent Caching
End-to-end QoS for Wired Applications
New challenges
Constantly changing system conditions
Network connectivity, user mobility
Device constraints
Energy, CPU, display, bandwidth
This needs
Provisioning, re-provisioning based on local conditions of wireless network
Data Placement based on current and projected data access patterns
Cross-layer awareness required
Resource provisioning algorithms utilize current system resource availability
information to ensure that applications meet their QoS requirements
QoS-based resource
provisioning
Provisioning Network and Server Resources Effectively
State information enables decision making for resource provisioning - e.g. Routing,
Scheduling and Placement
Maintaining accurate and current system information is important
Existing Approaches
In network : QoS Based Routing
At Server : Server Load Balancing
Future Middleware will support CPSS (Combined Path and Server
l
Scheduling) l : UF (l , r , n ), DL
l
l : BWavail , DLl
s1
s1 O s2
O CD
s3
s2 s
s : UF ( s1, r , n ), RSP
s3
Data Placement in Mobile
Environments
Initial placement
Design replication and intelligent
S1 S2 S3
data placement mechanisms that
Ensure effective resource v1 v2 v5 v6 v1 v2
management
v3 v4 v1 v2 v3 v4
Ensure QoS for admitted clients
Client
congestion
Distributed Adaptation
Distributed Adaptation
Cross-Layer Adaptation
Cross-Layer Adaptation
Power-Aware API
Appl. specific Adaptation
User/Application
Middleware
DYNAMO SYSTEM
Mohapatra et. al (ICN 05, ITCC 05,
ACM Middleware 04, DATE 04, ICDCS Operating System
03, MWCN 03, ACM MM 03, RTAS/RTSS
Workshops 03, Estimedia 03, CIPC 03, Architecture
ICDCS 01) (cpu, memory)
Network
Architecture
Cross-Layer Approach
device
User/Application
network
Distributed Middleware
Middleware
Operating System
Proxy Architecture
Quality/Power Matrix for COMPAQ IPAQ 3600 ( Grand Theft Auto Action Video Sequence)
Energy-Sensitive Video
Transcoding: Visual
Comparison
QoS
Monitor
NIC
Wireless
Network DISPLA
Y CPU
Proxy
Wireless Network
Mobile Device
Network Card Optimization
Wireless NIC cards can operate in various power modes
Avg. power consumption in sleep mode (0.184 W) whereas idle/receive modes
consume (1.34/1.435 W) respectively.
NIC can be transitioned to sleep mode (high energy savings)
Packets can get lost (quality drop)
Adaptation
Proxy buffers video and sends it in bursts to the device
Control info added - when device should wake up
Allows for long sleep intervals of the network card on device
Limitations
Large bursts can result in high packet loss rates
Access point and mobile device buffering limitations
Adaptation at Proxy
QSTREAM = Max(Qi) such that
PSTREAM * T < ERES
QSTREAM >
QTHRESHOLD
Video Stream (QSTREAM)
# of frames, buffer size, quality
Mobile
Proxy
Device
Residual Energy, Quality threshold
+ Noise (SL), Buffer capacity (Bf),
Decode rate (Fd)
QoS
NIC
adaptation
Monitor
E-Q profile
Network
Transcoder
Transmission Dynamo Middleware
Communication
Dynamo Middleware CPU OS
Linux NIC
NIC
Wireless
Network DISPLA
Y CPU
Proxy
Wireless Network
Mobile Device
Network Energy Savings NIC
Backlig
1738 Frames htCPU
Energy Savings 35% - 57% (over no optimization)
Frames Lost < 12%
399 Frames
Energy Savings 50% - 75% (over no optimization)
Frames Lost < 5%
2924 Frames
Energy Savings 25% - 45% (over no optimization)
Frames Lost < 10%
Backlight Compensation
Adaptation
1) Enhance luminance of video frames at proxy
Problems
Technique to achieve the suitable (optimal) backlight dimming factor
Reduce flicker induced by frequent backlight switching
Backlight Adaptation Backlight Quality Power
Level (PSNR) Savings
Backlight Power
Level Savings
Fair 80 44.8%
Good 125 39.7%
Excellent 186 30.3%
Backlight Power
Level Savings
QoS
NIC
adaptor
profileNetwork MonitorBacklight
E-Q adaptor
Transmission
B/L Scaling Transcoder Dynamo Middleware
Communication
Dynamo Middleware CPU OS
Linux NIC
NIC
Wireless
Network DISPLA
Y CPU
Proxy
Wireless Network
Mobile Device
Architecture/CPU
Adaptation
DVS idea: trade off processor speed for power
MPEG frame decoding good candidate
Frame decoding takes less than the frame delay
Decoding time depends on frame type: I, P, B
I B B P B B P
Voltage
no DVS
with DVS
MPEG Stream
0 D Fd time
CPU Adaptation using
Dynamo Video Stream (QSTREAM)
# of frames, buffer size,
quality
Backlight Setting + Profiled
WCET, BCET, Avg. Execution Mobile
Proxy
time Device
Residual Energy, Quality threshold
+ Noise (SL), Buffer capacity (Bf),
Decode rate (Fd)
Middleware communicates
Determine video quality execution characteristics to
Use a rule base of profiled scheduler
data to send execution Scheduler can now
parameters to the mobile dynamically re-compute
device slowdown parameters for the
new video quality
CPU Adaptation
User Negotiati
Profile on
Application
QoS CPU
NIC
adaptor adaptatio
Backlight n
Network Monitoradaptation
Rule BaseE-Q profileTransmission
B/L Scaling Transcoder Dynamo Middleware
Communication
Dynamo Middleware
schedule
r
Linux OS
NIC
Wireless
Network DISPLA
Y CPU
Proxy
Wireless Network
Mobile Device
Overall Energy Savings
Other OtherNetwork
(8.2%)
After Cross-Layer CPU (8.2%)(15.09%)
Network Savings
CPU
Optimizations (13.6%)
(37.7%)
(27.2%) NIC Savings
CPU (22.64%)
(13.6%)
Display Display Display
(26.9%) Savings(14.8%)
(12.1%)
Energy Distribution
Energy Distribution (After Optimization)
(Before Optimization) Note: These are avg. energy savings
Total energy savings ~ 48% (for a medium action video clip called Foreman)
Experimental Study
90 Negligible Energy Overhead
Measured Energy (Joules)
80
70 Encoding without Encryption
60
50 Encoding with Encryption
40 (Selective)
30
Encoding with Encryption
20
(Nave)
10
0
FOREMAN.qcif NEWS.qcif
Video Clips
Reliable Mobile Multimedia Energy
Implications
Problem and Motivation Idea: ME (Motion Estimation) is..
Experimental Study
Encoding Energy Consumption (iPAQ) PNSR Variation Encoding Energy Consumption (iPAQ)
25 36
25
30
20 20
Encoding Energy (J)
NO 24
Energy(J)
Encoding
PSNR (dB)
15
15 PBPAIR 18
PGOP-3 10
12
10 GOP-3 5
0.9
AIR-24 6
0.6
5 0
0
0.3 Intra_Th
0 0.2 0.4 0.6 0
0.8
1
11
13
15
17
19
21
25
27
39
41
43
45
47
49
23
29
31
33
35
37
1
0 Frame Number PLR
foreman akiyo garden PBPAIR PGOP-1 GOP-8 AIR-10 0-5 5-10 10-15 15-20 20-25
VS 2
VS 3
VS 1
Broker
Research Challenges
Resource Discovery
Application aspect
Ensure user QoS satisfaction
System View
Define an optimal grid resource allocation
User mobility
resource heterogeneity
Grid intermittent availability
Adapt to changing context
User mobility, device energy, and proxy availability
Data Placement
Application aspect
Guarantee data availability
Improve information retrieval efficiency with user mobility
System View
Address both short term on-demand caching and long term data placement
Make caching or data replication at finer granularity
Balance availability efficiency tradeoff
MapGrid
56
Contains information about 1. Resource Discovery
available resources on Grid 2. Schedule Planning
and Clients
Availabili ty ( j , t )
Gf j ( R, t )
LF j ( R, t ) Dist ( j , R) Distance of grid J from
request R.
58
Grid Resource Discovery for
Mobile Services
R: < IDR , T , QL , QH, , ER , itinerary>
Rescheduling if dynamic
changes happen
Grid Resource Discovery for
Mobile Services
Partition Service Period
N
All chunks have VS
Optimizing grid assignments?
resource selection
Factors: Load, location, availability, capacity
Achieve system-wide load balancing
Y
Maximize the number of accepted services
Graph theoretic approach
Rescheduling if dynamic
changes happen
Grid Service Discovery
(Multimedia Application)
Globus Toolkit
Information Broker
Services
Life time
JADE control
Http
Da
ta
The Volunteer Server can
join and disjoin the grid Client can specify
dynamically . QoS requirements via
friendly GUI
Client
JMF (Java Media Framework)
Mobile Cloud
Computing
What is Cloud Computing
Cloud computing is a model for enabling convenient, on-
demand network access to a shared pool of configurable
computing resources (e.g., networks, servers, storage,
applications, and services).
It can be rapidly provisioned and released with minimal
management effort.
It provides high level abstraction of computation and
storage model.
On-Demand Self Service,
Heterogeneous Access,
Resource Pooling.
Essential Characteristics
66
Essential Characteristics (cont.)
Resource Pooling:
The providers computing resources are pooled to serve multiple
consumers using a multi-tenant model.
Different physical and virtual resources dynamically assigned and
reassigned according to consumer demand.
Measured Service:
Cloud systems automatically control and optimize resources used
by leveraging a metering capability at some level of abstraction
appropriate to the type of service.
It will provide an analyzable and predictable computing platform.
67
Service Models
Cloud Software as a Service (SaaS):
The capability provided to the consumer is to use the providers
applications running on a cloud infrastructure.
The applications are accessible from various client devices such as
a web browser (e.g., web-based email).
The consumer does not manage or control the underlying cloud
infrastructure including network, servers, operating systems,
storage,
Examples: Caspio, Google Apps, Salesforce, Nivio, Learn.com.
68
Service Models (cont.)
69
Service Models (cont.)
70
Service Models (cont.)
71
Deployment Models
Private Cloud:
The cloud is operated solely for an organization. It may be managed by the
organization or a third party and may exist on premise or off premise.
Community Cloud:
The cloud infrastructure is shared by several organizations and supports a
specific community that has shared concerns.
It may be managed by the organizations or a third party and may exist on
premise or off premise.
Public Cloud:
The cloud infrastructure is made available to the general public or a large
industry group and it is owned by an organization selling cloud services.
Hybrid cloud:
The cloud infrastructure is a composition of two or more clouds (private,
community, or public).
72
Deployment Models (cont.)
73
Towards Pervasive Computing
Cloud
Computing
Evolution of Computing Environment: Promises
[Satyanarayanan_2001] :
Toward Pervasive Computing environment
74
Cloudlet
It has been shown that a one hop connection from mobile
device to internet is not efficient.
Humans are sensitive to the current delay in clouds
latency is unlikely to improve significantly
security and firewall it is unlikely that latency improves (although
increase in bandwidth).
Cloudlet
Between mobile devices and cloud pools (Infrastructure).
A cloudlet (micro edition of cloud)only
cloud contains soft
states such as cache copies of data or code.
75
Cloudlet (cont.)
Cloudlets are as the infrastructure for mobile cloud computing. From: [Satyanarayanan_2009]
76
Dynamic Task Reconfiguration
Solution
78
Source Parametric Flow Graph
Runtime on Device
Rd
ite
X1 Bd
Bi = Energy
i n X2
inf M1
A1 B1 cost of
Y1
A2 Xn
M2
B2
executing task
D P
device proxy Mi at the
An Y2
Mn
Bn
device.
Ai = Energy A p
Yn in
fi nite
cost of Rp
executing task Runtime on Proxy
Mi at the proxy.
If Mi is executing on the Device, then Xi = communication
costs in energy terms
If Mi is executing on the Proxy, then Yi = communication costs
79
High Level Algorithm
The minimum cut of the graph determines which
components can be moved to the proxy.
Can be solved using a modified FIFO Pre-Flow push
algorithm.
Complexity = O(n3)
81
AlfredO architecture and
applications (MW 2009)
Premise: Modularized application and Model of resource consumption and dependencies
among application modules
Goal: Partition the application between phone and server based on different criteria
(bandwidth, memory)
S7 S8 S5
Mobile Cut
Mobile Cut
S6 S1 S2 S3 S6 S1 S2 S3
S4 S4
S7 S8 S5 S7 S8 S5
83
AlfredO (MW 2009)
CHARISMA: Tiered Clouds
Just like ISPs cloud computing could be considered in different
Tiers.
The backbone Cloud which usually large organization such as
Google, Yahoo, Microsoft support them.
The second tier is edge Cloud which is near end user.
Qos
Qos Monitoring , Service Discovery
Monitoring
Qos Analyzer
Admission
Control
Scheduler Cloudlet Cloud
Pools Pools
Mobile
Client
Mobile Profile Monitoring
Mobile Profile Analyzer Analyzes mobile profile specially power level, location
pattern, for optimal scheduling.
Scheduler It schedules the accepted requests and design a plan
according to the negotiations done with Cloudlets (Schedule
Queue).
88
Cloudlet Cloud
89
Yahoo Cloud (Flicker) as a Cache.
Cloud(Yahoo)
Cloudlet
Cloud(Amazon)
90
http://www.youtube.com/watch?v=fQywFeN1wdM
91
Cloud computing
Remote access to distributed and shared cluster resources
Potentially owned by someone else (e.g. Amazon, Google, )
Users rent a small fraction of vast resource pools
Advertised service-level-agreements (SLAs)
Resources are opaque and isolated
Offer high availability, fault tolerance, and extreme scale
Relies on OS, network, and storage virtualization/isolation
Virtualization
SLAs
Web Services
Future: Interoperable
Networking Heterogeneous access
technologies available
802.11b
Wi-Fi
Bluetooth
Ad Hoc
Cellular
Applications are run over GPRS
......
QoS requirements
Non-predetermined traffic
flows start/terminate at Energy consumption
discrete time points. User/application
Access networks bandwidth specified access
and delay might fluctuate. preferences
Groupware apps
Multiparty collaborations
Seguti
Reliability + Security
Summary
Transparency
Mobile keeps its IP address when changing subnets
Communication continues after interruption of link possible
Compatibility
no changes to current end-systems and routers required - tunneling
no changes to fixed systems HA-FA tunneling
Security
authentication of all registration messages MIP messages
Efficiency and scalability
only a few additional messages to the mobile system required (connection typically via a
low bandwidth radio link) MIP messaging
Global support of a large number of mobile systems in the whole Internet HA for MN
router
end-system router
FA foreign
network
FA foreign
network
Ad-hoc network:
A collection of wireless mobile hosts forming a temporary
network without the aid of any established infrastructure
or centralized administration.
Significant differences to existing wired networks:
Wireless
Self-starting
No administrator
Cannot assume, that every computer is within
communication range of every other computer
Possibly quite dynamic topology of interconnections
Traffic types: unicast/multicast/anycast/geocast
Wireless Networking 112
Routing in MANET
Distributed
Bandwidth efficient
Reduce control traffic/overhead
Battery efficient
Fast route convergence
Correct: loop free
Reduce overhead
Unidirectional Link Support
Maximize
end-to-end throughput
delivery ratio
load balancing (congestion)
Minimize
end-to-end delay
packet loss
shortest path/minimum hop (route length)
overhead (bandwidth)
energy consumption
mobility slow (meeting) ~ fast (cars): slow (habitat) ~ fast: less focus on
focus on mobility mobility so far
source 2
source 1
source 2
Z
B
source 1
source 2
source 2
source 1
source 2
Z
source 1 & 2
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S E
F
B
C M L
J
A G
H D
K
I N
[S] Z
S E
F
B
C M L
J
A G
H D
K
I N
Z
S [S,E]
E
F
B
C M L
J
A [S,C] G
H D
K
I N
Z
S E
F [S,E,F]
B
C M L
J
A G
H D
[S,C,G] K
I N
Z
S E
F [S,E,F,J]
B
C M L
J
A G
H D
K
I [S,C,G,K] N
Z
S RREP [S,E,F,J,D]
E
F
B
C M L
J
A G
H D
K
I N
DATA [S,E,F,J,D] Z
S E
F
B
C M L
J
A G
H D
K
I N