Post on 26-Sep-2020
ACACIA – Context-aware Edge Computing for
Continuous Interactive Applications over
Mobile Networks
Junguk Cho, Jacobus Van der Merwe,
Karthikeyan Sundaresan, Rajesh Mahindra,
and Sampath Rangarajan
1
• Emerging Continuous Interactive (CI) mobile applications at scale
Continuous Interactive (CI) Mobile Apps
2
• Emerging Continuous Interactive (CI) mobile applications at scale• Example of Continuous Interactive (CI) mobile apps
Continuous Interactive (CI) Mobile Apps
3
• Emerging Continuous Interactive (CI) mobile applications at scale• Example of Continuous Interactive (CI) mobile apps
• Augmented reality (AR)
Continuous Interactive (CI) Mobile Apps
4
• Emerging Continuous Interactive (CI) mobile applications at scale• Example of Continuous Interactive (CI) mobile apps
• Augmented reality (AR) • Face Recognition
Continuous Interactive (CI) Mobile Apps
5
• Emerging Continuous Interactive (CI) mobile applications at scale• Example of Continuous Interactive (CI) mobile apps
• Augmented reality (AR) • Face Recognition• Virtual Reality (VR)
Continuous Interactive (CI) Mobile Apps
6
• Emerging Continuous Interactive (CI) mobile applications at scale• Example of Continuous Interactive (CI) mobile apps
• Augmented reality (AR) • Face Recognition• Virtual Reality (VR)• Autonomous driving
Continuous Interactive (CI) Mobile Apps
7
Characteristics of CI Mobile Apps• Highly responsive (~ 100 ms)
• Overlay information (graphics, text or video) should be shown in real-time
• Intensive computation • Too slow to run completely on mobile devices
• Have centralized databases in the server
8
Enablers for CI Mobile Apps
9
Computation offloading to cloud
Enablers for CI Mobile Apps
10
Mobile network
Enablers for CI Mobile Apps
11
Enablers for CI Mobile Apps
12
User Context
Enablers for CI Mobile Apps
13
Each Components Have Evolved a Lot
14
Standalone Approach Is Not Sufficient• The end to end latency of CI apps is affected by various factors
• Network conditions (bandwidth & latency)• Application computation latency (object matching computation)
15
Standalone Approach Is Not Sufficient• The end to end latency of CI apps is affected by various factors
• Network conditions (bandwidth & latency)• Application computation latency (object matching computation)
• Each approach treats others as blackbox• Computation offloading framework
• No consideration of mobile network complexities
• Evolving mobile network & context information• No well-defined protocol and interface to make synergies among them
16
?
Need a General and Holistic E2E Approach
Application
Mobile Network
User Context
17
Key Questions?
18
How should the three entities (application, network, and user) be jointly orchestrated and combined in a service abstraction over mobile networks to enable CI mobile applications?
?
Application
Mobile Network
User Context
ACACIA
•A service abstraction frameworks could be provided by mobile operators
19
ACACIA
•A service abstraction frameworks could be provided by mobile operators
•A general and holistic end-to-end approach to enabling CI services on edge clouds over mobile network
20
ACACIA
•A service abstraction frameworks could be provided by mobile operators
•A general and holistic end-to-end approach to enabling CI services on edge clouds over mobile network
• Leverage client context information through LTE-direct• Orchestrate three entities • Optimize both network and application processing
21
ACACIA Usecase – Retail Mall
22
AR Available in Laptop Section
Laptop
SALESMAN
ACACIA Usecase – Retail Mall
AR Available in Food Section
Food SALESMAN
23
AR Available in Laptop Section
Laptop
SALESMAN
ACACIA Usecase – Retail Mall
AR Available in Food Section
Food SALESMAN
24
Retail App
“Interest” in laptop
Laptop
SALESMAN
ACACIA Usecase – Retail Mall
Food SALESMAN
Retail App
“Interest” in laptop
“Interest” Match
“Interest” non- Match
25
AR Available in Laptop Section
AR Available in Food Section
Laptop
SALESMAN
ACACIA Usecase – Retail Mall
Food SALESMAN
Retail AR App
“Interest” in laptop
Mobile Edge Clouds
Retail AR server
26
AR Available in Laptop Section
AR Available in Food Section
ACACIA Architecture
27
User Context Discovery
28
ACACIA Mobile Edge Network
29
Context-aware Application Optimization
30
User Context Discovery
31
Android Framework
SUBSCRIBER
Modem(LTE)
Sub/CI
App
ACACIA
device
manager
• Works as a proxy between CI apps and LTE modem
ACACIA Device Manager
32
Push“Interest”
DISCOVERYSUBSCRIBE
FILTER
Android Framework
PUBLISHER
Modem(LTE)
Pub App
ACACIA
device
manager
DISCOVERYPUBLISH FILTER
Android Framework
SUBSCRIBER
Modem(LTE)
Sub/CI
App
ACACIA
device
manager
• Works as a proxy between CI apps and LTE modem
ACACIA Device Manager
33
DISCOVERYSUBSCRIBE
FILTER
Android Framework
PUBLISHER
Modem(LTE)
Pub App
ACACIA
device
manager
DISCOVERYPUBLISH FILTER
Listen
Android Framework
SUBSCRIBER
Modem(LTE)
Sub/CI
App
ACACIA
device
manager
• Works as a proxy between CI apps and LTE modem
ACACIA Device Manager
34
DISCOVERYSUBSCRIBE
FILTER
Android Framework
PUBLISHER
Modem(LTE)
Pub App
ACACIA
device
manager
DISCOVERYPUBLISH FILTER
Listen
Push “Service”
Android Framework
SUBSCRIBER
Modem(LTE)
Sub/CI
App
ACACIA
device
manager
• Works as a proxy between CI apps and LTE modem
ACACIA Device Manager
35
DISCOVERYSUBSCRIBE
FILTER
Android Framework
PUBLISHER
Modem(LTE)
Pub App
ACACIA
device
manager
DISCOVERYPUBLISH FILTER
Listen LTE-DirectBROADCAST
Android Framework
SUBSCRIBER
Modem(LTE)
Sub/CI
App
ACACIA
device
manager
• Works as a proxy between CI apps and LTE modem
ACACIA Device Manager
36
DISCOVERYSUBSCRIBE
FILTER
Android Framework
PUBLISHER
Modem(LTE)
Pub App
ACACIA
device
manager
DISCOVERYPUBLISH FILTER
LTE-DirectBROADCAST
IF MATCH ?
• Manage the network connectivity on demand
ACACIA Device Manager
37
First Match
• Manage the network connectivity on demand
ACACIA Device Manager
38
First Match
39
ACACIA Mobile Edge Network
LTE GW-U
Internet
Mobile Edge Clouds
CI Server
Local LTE GW-U
40
ACACIA Mobile Edge Network
Mobile CoreComponent
eNodeB
LTE GW-U
Internet
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
41
ACACIA Mobile Edge Network
Mobile CoreComponent
eNodeB
LTE GW-U
Internet
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
42
ACACIA Mobile Edge Network
Mobile CoreComponent
eNodeB
LTE GW-U
Internet
Mobile Edge Clouds
CI Server
Local LTE GW-U
43
ACACIA Mobile Edge Network
Mobile CoreComponent
eNodeB
LTE GW-U
Internet
Mobile Edge Clouds
CI Server
Local LTE GW-U
44
ACACIA Mobile Edge Network
Mobile CoreComponent
eNodeB
• On demand connectivity• Reduce control overhead due to two always-on connectivity
• Fine-grained CI traffic control in LTE eNodeB and LTE GWs based on service types
• No Impact on the rest of traffic
• Cost effective • Without deploying middlebox for selective CI traffic
• 3GPP standard compatible • Without modification of eNodeB, LTE interfaces and protocol
Requirements of ACACIA Mobile Edge Network
45
• On demand connectivity• Reduce control overhead due to two always-on connectivity
• Fine-grained CI traffic control in LTE eNodeB and LTE GWs based on service types
• No Impact on the rest of traffic
• Cost effective • Without deploying middlebox for selective CI traffic
• 3GPP standard compatible • Without modification of eNodeB, LTE interfaces and protocol
Requirements of ACACIA Mobile Edge Network
46
LTE/EPC QoS bearer framework with LTE Gateways using SDN & NFV
• Request mobile edge clouds connectivity
47
Steps to Set up Mobile Edge Connectivity
LTE GW-U
Internet
ACACIADeviceManager
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
Mobile CoreComponent
eNodeBFirst Match
• LTE/EPC QoS bearer framework
48
Steps to Set up Mobile Edge Connectivity
LTE GW-U
Internet
ACACIADeviceManager
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
Mobile CoreComponent
eNodeB
• LTE/EPC QoS bearer framework
49
Steps to Set up Mobile Edge Connectivity
LTE GW-U
Internet
ACACIADeviceManager
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
Mobile CoreComponent
eNodeB
• LTE Gateways using SDN & NFV
50
Steps to Set up Mobile Edge Connectivity
LTE GW-U
Internet
ACACIADeviceManager
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
Mobile CoreComponent
eNodeB
• LTE Gateways using SDN & NFV
51
Steps to Set up Mobile Edge Connectivity
LTE GW-U
Internet
ACACIADeviceManager
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
Mobile CoreComponent
eNodeB
• LTE Gateways using SDN & NFV
52
Steps to Set up Mobile Edge Connectivity
LTE GW-U
Internet
ACACIADeviceManager
Mobile Edge Clouds
CI Server
CI APP
Local LTE GW-U
Mobile CoreComponent
eNodeB
Traffic classification in LTE modembased on 5 tuples or other info
ServiceInterests
53
?
?
ServiceInterests
54
Mobile Edge Network
> 1 sec
0
0.2
0.4
0.6
0.8
1
1.2
Ob
ject
Mat
chin
g Ti
me
(se
c)
1440 * 1080 Image
1 Object 5 Objects 10 Objects 25 Objects 50 Objects
Application Processing is Still Heavy
55
Database pruning!
0
0.2
0.4
0.6
0.8
1
1.2
Ob
ject
Mat
chin
g Ti
me
(se
c)
1440 * 1080 Image
1 Object 5 Objects 10 Objects 25 Objects 50 Objects
Application Optimization
56
Context-aware Application Optimization
57
User Location Context
58
Geo-tagged
AR Database
Section 0 User Loc
Searching space
(section 6)
User Location Context
59
• Use standard trilateration localization
Geo-tagged
AR Database
Section 0 User Loc
Searching space
(section 6)
Publishers as Landmarks
60
• Use standard trilateration localization• Publishers act as landmarks• Use rxPower values from every service discovery message
Geo-tagged
AR Database
Publishers
(Landmarks)Section 0
Context-aware Application Optimization
61
AR
Front-end
Mobile device
ACACIA
device manager
CI Server on Edge Cloud
(Trilateration solver)
LTE-direct
Localization Manager
Matcher
Frames
(rxPower,
Landmark nam
e)
(Discovery msg,
rxPower)
User Context Data AR Data
AR
Back-end
Context-aware Application Optimization
62
AR
Front-end
Mobile device
ACACIA
device manager
CI Server on Edge Cloud
(Trilateration solver)
LTE-direct
Localization Manager
Matcher
Frames
(rxPower,
Landmark nam
e)
(Discovery msg,
rxPower)
User Context Data AR Data
AR
Back-end
Context-aware Application Optimization
63
AR
Front-end
Mobile device
ACACIA
device manager
CI Server on Edge Cloud
(Trilateration solver)
LTE-direct
Localization Manager
Matcher
Frames
(rxPower,
Landmark nam
e)
(Discovery msg,
rxPower)
User Context Data AR Data
AR
Back-endDB
Get images
for cell 6
Searching
space (cell 6)
LandmarksCell 0
Context-aware Application Optimization
64
AR
Front-end
Mobile device
ACACIA
device manager
CI Server on Edge Cloud
(Trilateration solver)
LTE-direct
Localization Manager
Matcher
Frames
(rxPower,
Landmark nam
e)
(Discovery msg,
rxPower)
User Context Data AR Data
AR
Back-endDB
Get images
for cell 6
Searching
space (cell 6)
LandmarksCell 0
Context-aware Application Optimization
65
AR
Front-end
Mobile device
ACACIA
device manager
CI Server on Edge Cloud
(Trilateration solver)
LTE-direct
Localization Manager
Matcher
Frames
(rxPower,
Landmark nam
e)
(Discovery msg,
rxPower)
User Context Data AR Data
AR
Back-endDB
Get images
for cell 6
Searching
space (cell 6)
LandmarksCell 0
66
LocationDatabase pruning
?
67
ServiceInterestsLocation
Database pruning
Mobile Edge Network
• Use OpenEPC for LTE core network components (MME, PCRF, PCEF)
• ACACIA mobile edge network• Extend OpenEPC to support split LTE Gateways and QoS framework• Use Open vSwitch and Ryu SDN controller for Local LTE gateways
• ACACIA Device Manager• Implement it as Android Service using Messenger class in android
• AR-based Retail Application• Pub-Sub GUI application• Use OpenCV library (SURF) for object matching• Geo-tagged object database• Trilateration localization solver
ACACIA Implementation
68
LTE IP.access (small cell)
One+ One
OpenEPC Core Network+ GW-Us + MEC server
LTE Basestation
OpenEPC Core Network
GW-Us
MEC
MEC server
One+ One
http://phantomnet.org/ACACIA Evaluation
69
• Microbenchmark• ACACIA standard compliance• LTE GW performance• LTE-direct localization accuracy
• Impact of network optimization• Impact of application optimization• End-to-end evaluation
ACACIA Evaluation Criteria
70
• Microbenchmark• ACACIA standard compliance• LTE GW performance• LTE-direct localization accuracy
• Impact of network optimization• Impact of application optimization• End-to-end evaluation
ACACIA Evaluation Criteria
71
Benchmark : AR application with geo-tagged AR DB ( 105 objects in 21 sections)
End-to-end Evaluation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Match Compute Network Total
End
-to
-En
d L
ate
ncy
(se
c)
ACACIA MEC CLOUD
72
End-to-end Evaluation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Match Compute Network Total
End
-to
-En
d L
ate
ncy
(se
c)
ACACIA MEC CLOUD
73
Benchmark : AR application with geo-tagged AR DB ( 105 objects in 21 sections)
End-to-end Evaluation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Match Compute Network Total
End
-to
-En
d L
ate
ncy
(se
c)
ACACIA MEC CLOUD
74
Benchmark : AR application with geo-tagged AR DB ( 105 objects in 21 sections)
End-to-end Evaluation
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Match Compute Network Total
End
-to
-En
d L
ate
ncy
(se
c)
ACACIA MEC CLOUD
ACACIA << MEC << CLOUD 75
Benchmark : AR application with geo-tagged AR DB ( 105 objects in 21 sections)
• Propose ACACIA - a service abstraction framework to enable CI apps
• Give insights to design mobile edge computing and usecase of LTE-direct on 5G network
• Validate ACACIA design with smartphones, LTE base station and Software-based EPC
https://wiki.phantomnet.org/wiki/phantomnet/acacia
Conclusion
76