Introductory presentation for Ph.D. thesis of Nitinder Mohan titled "Edge Computing Platforms and Protocols". The defense took place at the University of Helsinki, Finland on 8th November 2019.
The video of the presentation is available at https://youtu.be/dDVZozTwreE
The thesis can be found on https://helda.helsinki.fi/handle/10138/306041
2. Cloud Computing
• Global network of inter-connected datacenters
• Datacenters managed and operated by cloud
providers, e.g. Google, Microsoft, Amazon etc.
• Developers can deploy their applications on
virtualized resources of the cloud worldwide
1
More than 60% of Internet workloads are cloud-based
3. Next Generation Applications
Trends by year 2025
• More than 75 billion connected devices
• Expected 11 trillion USD market share
• On average 7 sensors per person worldwide
Requirements
1. Latency-critical processing
2. Big data aggregation and analysis
3. Location and context aware computations
Internet-of-Things
3
4. Problem: Network!
Ø High transport cost
Ø High data volume
Ø High network latency
Can Cloud support next-generation applications?
4
5. Edge Computing
Small-scale server(s) deployed near the users to
compute generated data
Benefits:
üDecreased latency for computation
üReduced network load due to pre-processing
üLocation and contextual awareness
Network
DatacenterEdge
Server
User
5
6. Problems in Edge computing
1. Unmanaged hardware
2. Constrained network
3. Inconsistent reliability and availability
4. Lack of standardized protocols
6
8. Hardware
Deployment, Cooling, Maintenance
Infrastructure
Processing, Storage, Networking
Platform
Virtual Machines, Containers,
Databases
Software
Cloud Computing Service Model Edge Computing Service Model
Hardware
Organization? Installation?
Maintenance? Security?
Infrastructure
Mobile hardware? Limited
processing? Low storage?
Wireless/Cellular NICs?
Platform
Server Discovery? Self-Organizing
Containers?
Parental Control, Firewall, Content
Catalog, User Authentication
Factory Automation, Intelligence,
Autonomous vehicles, IoT Analytics
Cloud providers?
Crowdsourced?
ISPs?
Light-weight VM?
Mirage OS?
Unikernels?
Smartphone
manufactures?
Specialized
servers?
7
9. Hardware Thesis Contributions
Edge-Fog Cloud: All-inclusive, node-oriented edge computing architecture which logically
categorizes resources in layers.
Anveshak: Deployment framework that assists service providers to identify best locations in a
geographical region for installing edge servers.
Infrastructure
QAware: Cross-layer scheduler for Multipath TCP which allows edge servers to use multiple
network paths simultaneously while overcoming excessive buffering and delays on any path.
Data Grouping protocol: Edge caching mechanism which predicts and pre-caches prerequisite
data in local caches of edge servers for upcoming computations.
LPCF & eLPCF: Task allocation frameworks which distributes application jobs on a cluster of
edge servers while minimizing processing, networking and energy costs.
Platform
ICON: Self-managing virtualized containers which can automatically migrate and replicate to servers
experiencing more user traffic without involving application owner.
ExEC: Open platform which allows cloud providers to discover and utilize resources offered by third-
party edge providers operating in the network.
8
10. Hardware Thesis Contributions
Edge-Fog Cloud: All-inclusive, node-oriented edge computing architecture which logically
categorizes resources in layers.
Anveshak: Deployment framework that assists service providers to identify best locations in a
geographical region for installing edge servers.
Infrastructure
QAware: Cross-layer scheduler for Multipath TCP which allows edge servers to use multiple
network paths simultaneously while overcoming excessive buffering and delays on any path.
Data Grouping protocol: Edge caching mechanism which predicts and pre-caches prerequisite
data in local caches of edge servers for upcoming computations.
LPCF & eLPCF: Task allocation frameworks which distributes application jobs on a cluster of
edge servers while minimizing processing, networking and energy costs.
Platform
ICON: Self-managing virtualized containers which can automatically migrate and replicate to servers
experiencing more user traffic without involving application owner.
ExEC: Open platform which allows cloud providers to discover and utilize resources offered by third-
party edge providers operating in the network.
8
11. Hardware Thesis Contributions
Edge-Fog Cloud: All-inclusive, node-oriented edge computing architecture which logically
categorizes resources in layers.
Anveshak: Deployment framework that assists service providers to identify best locations in a
geographical region for installing edge servers.
Infrastructure
QAware: Cross-layer scheduler for Multipath TCP which allows edge servers to use multiple
network paths simultaneously while overcoming excessive buffering and delays on any path.
Data Grouping protocol: Edge caching mechanism which predicts and pre-caches prerequisite
data in local caches of edge servers for upcoming computations.
LPCF & eLPCF: Task allocation frameworks which distributes application jobs on a cluster of
edge servers while minimizing processing, networking and energy costs.
Platform
ICON: Self-managing virtualized containers which can automatically migrate and replicate to servers
experiencing more user traffic without involving application owner.
ExEC: Open platform which allows cloud providers to discover and utilize resources offered by third-
party edge providers operating in the network.
How to do computations on edge clouds?
12. Hardware Thesis Contributions
Edge-Fog Cloud: All-inclusive, node-oriented edge computing architecture which logically
categorizes resources in layers.
Anveshak: Deployment framework that assists service providers to identify best locations in a
geographical region for installing edge servers.
Infrastructure
QAware: Cross-layer scheduler for Multipath TCP which allows edge servers to use multiple
network paths simultaneously while overcoming excessive buffering and delays on any path.
Data Grouping protocol: Edge caching mechanism which predicts and pre-caches prerequisite
data in local caches of edge servers for upcoming computations.
LPCF & eLPCF: Task allocation frameworks which distributes application jobs on a cluster of
edge servers while minimizing processing, networking and energy costs.
Platform
ICON: Self-managing virtualized containers which can automatically migrate and replicate to servers
experiencing more user traffic without involving application owner.
ExEC: Open platform which allows cloud providers to discover and utilize resources offered by third-
party edge providers operating in the network.
8
13. Requirement of Edge Computing
Need of a reference architecture for designing protocols and
platforms!
Constraints? Capabilities? Capacities?
9
14. Edge-Fog Cloud (EFCloud)
Edge
Ø Collection of devices:
i. Loosely-coupled
ii. Voluntary/Crowdsourced
iii. Owned by independent operators
Ø Extremely close to sensors & clients
Ø Device-to-device connectivity over
WiFi/cellular/Bluetooth/Zigbee etc.
Ø Varying processing capability
e.g. desktops, laptops, workstations,
nano data centers etc. 10
15. Edge-Fog Cloud (EFCloud)
Fog
Ø High capacity compute servers
Ø Co-located with networking devices
Ø Designed, manufactured, managed
and deployed by cloud vendors
Ø Lies farther from sensors and clients
Ø Dense connectivity within layer
Ø Support for virtualization
technologies
e.g. routers, switches, basestations etc.
11
16. Edge-Fog Cloud (EFCloud)
Data Store
Ø Data archival and storage
Ø No computation on data
Ø Reliable, ease-of-access, secure and
global access to storage for Edge and
Fog resources
12
17. Edge-Fog Cloud (EFCloud)
Salient Features
1. All-inclusive edge architecture
2. Native support for mobility
3. Satisfies emerging application
requirements
4. Context-aware computation
5. No single point-of-failure
6. No vendor lock-in
13
18. Hardware Thesis Contributions
Edge-Fog Cloud: All-inclusive, node-oriented edge computing architecture which logically
categorizes resources in layers.
Anveshak: Deployment framework that assists service providers to identify best locations in a
geographical region for installing edge servers.
Infrastructure
QAware: Cross-layer scheduler for Multipath TCP which allows edge servers to use multiple
network paths simultaneously while overcoming excessive buffering and delays on any path.
Data Grouping protocol: Edge caching mechanism which predicts and pre-caches prerequisite
data in local caches of edge servers for upcoming computations.
LPCF & eLPCF: Task allocation frameworks which distributes application jobs on a cluster of
edge servers while minimizing processing, networking and energy costs.
Platform
ICON: Self-managing virtualized containers which can automatically migrate and replicate to servers
experiencing more user traffic without involving application owner.
ExEC: Open platform which allows cloud providers to discover and utilize resources offered by third-
party edge providers operating in the network.
14
30. Least Processing Cost First (LPCF)
I. Optimize Processing Time
Minimize:
Linear Assignment Problem
$
',)∈,
𝐶
𝐽7'89(𝑖)
𝐷:;/.(𝑗)
𝑥')
3 2 2 5 6
4 2 5 4 2
Dproc [i] =
Jsize [i] =
20
31. I. Optimize Processing Time
Minimize:
Linear Assignment Problem
• Solved using Kuhn-Munkres/
Hungarian algorithm
• Optimal solution guaranteed in
O(n3)
Least Processing Cost First (LPCF)
D1:3 D2:2 D3:2
D4:5 D5:6
1
4 34
1
J1:4 J2:2 J5:2
J4:4 J3:5
Least Processing Cost: 4.966
$
',)∈,
𝐶
𝐽7'89(𝑖)
𝐷:;/.(𝑗)
𝑥')
20
32. Least Processing Cost First (LPCF)
II. Create sub-problem space
Interchange homogeneous devices
running homogeneous jobs
Search Space Calculation:
1. Same processing power
→ interchange jobs
2. Same job size
→ interchange devices
D1:3 D2:2 D3:2
D4:5 D5:6
1
4 34
1
J1:4 J2:2 J5:2
J4:4 J3:5
J1:4 J5:2 J2:2
J4:4 J3:5
J4:4 J5:2 J2:2
J1:4 J3:5
21
33. Least Processing Cost First (LPCF)
II. Create sub-problem space
Interchange homogeneous devices
running homogeneous jobs
Search Space Calculation:
1. Same processing power
→ interchange jobs
2. Same job size
→ interchange devices
Least Processing Cost: 4.966
D1 D2 D3 D4 D5
1. J1 J2 J5 J4 J3
2. J1 J5 J2 J4 J3
3. J4 J5 J2 J1 J3
4. J4 J2 J5 J1 J3
21
34. Least Processing Cost First (LPCF)
III. Least Network Cost
1. Compute network cost of each
assignment
2. Choose the assignment with
least network cost
𝐽./00 𝑖, 𝑗 ∗ 𝐷./00(𝑓 𝑖 , 𝑓(𝑗))
D1 D2 D3 D4 D5
1. J1 J2 J5 J4 J3
2. J1 J5 J2 J4 J3
3. J4 J5 J2 J1 J3
4. J4 J2 J5 J1 J3
N/W
20
27
19
28
Least Processing Cost: 4.966
22
35. Performance Overview
5 20 40 60 80 100
Topology Size
0
1000
2000
3000
4000
NetworkCost
NOC
LPCF
Minimum bound
Maximum bound
20 40 60 80 100
Topology Size
0
50
100
150
200
250
ProcessingCost
NOC
LPCF
eLPCF
5 20 40 60 80 100
Topology Size
5
10
15
20
25
30
%decreaseinenergy
Edge
Fog
Network cost remains
within 10% range of
the optimal value
Processing cost is
multiple times lower
than de-facto solvers
10% decrease in overall energy
used. More than 20% energy
saved for battery constrained
devices
24
36. Hardware Thesis Contributions
Edge-Fog Cloud: All-inclusive, node-oriented edge computing architecture which logically
categorizes resources in layers.
Anveshak: Deployment framework that assists service providers to identify best locations in a
geographical region for installing edge servers.
Infrastructure
QAware: Cross-layer scheduler for Multipath TCP which allows edge servers to use multiple
network paths simultaneously while overcoming excessive buffering and delays on any path.
Data Grouping protocol: Edge caching mechanism which predicts and pre-caches prerequisite
data in local caches of edge servers for upcoming computations.
LPCF & eLPCF: Task allocation frameworks which distributes application jobs on a cluster of
edge servers while minimizing processing, networking and energy costs.
Platform
ICON: Self-managing virtualized containers which can automatically migrate and replicate to servers
experiencing more user traffic without involving application owner.
ExEC: Open platform which allows cloud providers to discover and utilize resources offered by third-
party edge providers operating in the network.
25
37. Hardware PublicationsInfrastructurePlatform
Anveshak: Placing Edge Servers In The Wild. N. Mohan, A. Zavodovski, P. Zhou, and J. Kangasharju, MECOMM 2018
Placing it right!: optimizing energy, processing, and transport in Edge-Fog clouds. N. Mohan and J. Kangasharju, Annals of
Telecommunication 2018
Managing Data in Computational Edge Cloud. N. Mohan, P. Zhou, K. Govindaraj and J. Kangasharju, MECOMM 2017
QAware: A Cross-Layer Approach to MPTCP Scheduling. T. Shreedhar, N. Mohan, S.K. Kaul, & J. Kangasharju. IFIP Networking 2018
ExEC: Elastic Extensible Edge Cloud. A. Zavodovski, N. Mohan, S. Bayhan, W. Wong and J. Kangasharju, EdgeSys 2019
ICON: Intelligent Container Overlays. A. Zavodovski, N. Mohan, S. Bayhan, W. Wong and J. Kangasharju, HotNets 2018
Is two greater than one?: Analyzing Multipath TCP over Dual-LTE in the Wild. N. Mohan, T. Shreedhar, A. Zavodovski, J. Kangasharju
& S.K. Kaul. Manuscript 2019.
25
42. Edge computing
Multiple edge computing approaches have
been proposed
i. Cloudlets → Miniature datacenters
Network
43. Edge computing
Multiple edge computing approaches have
been proposed
i. Cloudlets → Miniature datacenters
ii. Fog → Network devices (Base stations, Routers)
Network
44. Edge computing
Multiple edge computing approaches have
been proposed
i. Cloudlets → Miniature datacenters
ii. Fog → Network devices (Base stations, Routers)
iii. Edge → Crowdsourced (smartphones, smart
speakers)
Network
45. Edge computing
Multiple edge computing approaches have
been proposed
i. Cloudlets → Miniature datacenters
ii. Fog → Network devices (Base stations, Routers)
iii. Edge → Crowdsourced (smartphones, smart
speakers, ..)
iv. Mist → Data generators (sensors,
microcontrollers, ..)
Network
47. Hardware
Deployment, Cooling, Maintenance
Infrastructure
Processing, Storage, Networking
Platform
Virtual Machines, Containers,
Databases
Software
Cloud Computing Service Model Edge Computing Service Model
Parental Control, Firewall, Content
Catalog, User Authentication
Factory Automation, Intelligence,
Autonomous vehicles, IoT Analytics
48. Hardware
Deployment, Cooling, Maintenance
Infrastructure
Processing, Storage, Networking
Platform
Virtual Machines, Containers,
Databases
Software
Cloud Computing Service Model Edge Computing Service Model
Hardware
Organization? Installation?
Maintenance? Security?
Parental Control, Firewall, Content
Catalog, User Authentication
Factory Automation, Intelligence,
Autonomous vehicles, IoT Analytics
Cloud providers?
Crowdsourced?
Cellular providers?
49. Hardware
Deployment, Cooling, Maintenance
Infrastructure
Processing, Storage, Networking
Platform
Virtual Machines, Containers,
Databases
Software
Cloud Computing Service Model Edge Computing Service Model
Hardware
Organization? Installation?
Maintenance? Security?
Infrastructure
Mobile hardware? Limited
processing? Low storage?
Wireless/Cellular NICs?
Parental Control, Firewall, Content
Catalog, User Authentication
Factory Automation, Intelligence,
Autonomous vehicles, IoT Analytics
Cloud providers?
Crowdsourced?
Cellular providers?
50. Hardware
Deployment, Cooling, Maintenance
Infrastructure
Processing, Storage, Networking
Platform
Virtual Machines, Containers,
Databases
Cloud Computing Service Model Edge Computing Service Model
Hardware
Organization? Installation?
Maintenance? Security?
Infrastructure
Mobile hardware? Limited processing?
Low storage? Wireless/Cellular NICs?
Platform
Server Discovery? Self-Organizing
Containers? Unikernels?
Cloud providers?
Crowdsourced?
Cellular providers?
Software
Parental Control, Firewall, Content
Catalog, User Authentication
Factory Automation, Intelligence,
Autonomous vehicles, IoT Analytics
51. Hardware Thesis Research Questions
RQ2: Can independent entities enroll their compute resources in an existing edge cloud
platform?
RQ1: Where should the cloud providers install compute servers in the physical world to
satisfy the application requirements at the "edge"?
Infrastructure
RQ6: How do we assure datacenter-like network behavior over edge servers which operate
on a public wireless network?
RQ5: Can existing network technologies available at the edge support the requirements
imposed by end-applications for optimal performance?
RQ4: How to pre-cache computational data within edge servers to improve computations?
RQ3: How do we utilize availability and variability of edge servers for computing tasks?
Platform
RQ9: How can independent edge providers generate revenue at par with cloud providers?
RQ8: How can edge cloud ensure the promised Quality-of-Service despite variability in user
requests and infrastructure hardware?
RQ7: How can existing cloud virtualization technologies be exploited in edge clouds?
52. Hardware Thesis Research Questions
RQ2: Can independent entities enroll their compute resources in an existing edge cloud
platform?
RQ1: Where should the cloud providers install compute servers in the physical world to
satisfy the application requirements at the "edge"?
Infrastructure
RQ6: How do we assure datacenter-like network behavior over edge servers which operate
on a public wireless network?
RQ5: Can existing network technologies available at the edge support the requirements
imposed by end-applications for optimal performance?
RQ4: How to pre-cache computational data within edge servers to improve computations?
RQ3: How do we utilize availability and variability of edge servers for computing tasks?
Platform
RQ9: How can independent edge providers generate revenue at par with cloud providers?
RQ8: How can edge cloud ensure the promised Quality-of-Service despite variability in user
requests and infrastructure hardware?
RQ7: How can existing cloud virtualization technologies be exploited in edge clouds?
53. Can independent entities enroll their compute
resources in an existing edge cloud platform?
Hardware
Infrastructure
Platform
55. Deploying Tasks on Edge Clouds
Edge clouds cannot use existing datacenter-based assignment
protocols as Edge and Fog recourses can:
1. operate in non-to-semi unmanaged environments with variable
availability
2. be equipped with constrained processing hardware requiring
multiple servers to complete single task
3. be inter-connected via wireless network links prone to latency
and congestion
4. be powered by limited battery capacity
56. Task Deployment on EFCloud
Network-Only Cost Solver
1. Finds job placement which has
the least possible network cost
2. Very large problem search space
3. NP-hard optimization with no
guarantee for optimal solution
Least Processing Cost First
1. Finds job placement with least
processing cost and almost-least
networking cost
2. Reduced problem search space
3. Placement guaranteed for any
problem size in linear time
23
58. Deploying Cloud Applications
Cloud applications are composed of
multiple integrated (micro)services
Catalog Retrieval
Parental Control
Application Metrics
Exception Tracking
Audit Logging
Health Check
…
700+
59. Deploying Cloud Applications
Cloud applications are composed of
multiple integrated (micro)services
→ Every microservice is encapsulated
as virtualized container which is
then deployed on the cloud
Parental Control
Catalog Retrieval
Catalog Retrieval
Parental Control
Application Metrics
Exception Tracking
Audit Logging
Health Check
…
700+
65. Fragmentation in EFCloud
EFCloud is amalgamation of different managing
entities with strict authoritative boundaries
1. ISP-backed Fog + Cellular-based Edge
EdgeFog
66. Fragmentation in EFCloud
EFCloud is amalgamation of different managing
entities with strict authoritative boundaries
1. ISP-backed Fog + Cellular-based Edge
2. WiFi router Fog + Crowdsourced Edge
EdgeFog
67. Fragmentation in EFCloud
EFCloud is amalgamation of different managing
entities with strict authoritative boundaries
1. ISP-backed Fog + Cellular-based Edge
2. WiFi router Fog + Crowdsourced Edge
3. Fog of cloud providers + Independent Edge
servers
EdgeFog
68. Fragmentation in EFCloud
EFCloud is amalgamation of different managing
entities with strict authoritative boundaries
1. ISP-backed Fog + Cellular-based Edge
2. WiFi router Fog + Crowdsourced Edge
3. Fog of cloud providers + Independent Edge
servers
Multiple entities compete in/across every
grouping for larger market share
→ Heterogeneity in Hardware and Software!
EdgeFog
70. Intelligent Containers (ICONs)
Self-managing virtualized services which can intelligently
and automatically move to edge servers* nearest to
incoming user requests
*irrespective of who owns/operates them
71. Operation of ICON
ICON Edge/Fog
End-users
I. Initially, ICON is in the cloud
One or multiple origination points
72. Operation of ICON
ICON
End-users
Edge/Fog
I. Initially, ICON is in the cloud
One or multiple origination points
II. ICON monitors incoming flows
Where user requests are coming from?
73. Operation of ICON
ICON
End-users
Discover
Edge/Fog
I. Initially, ICON is in the cloud
One or multiple origination points
II. ICON monitors incoming flows
Where user requests are coming from?
III. ICON discovers deployment locations
In the domain of end-users or on a path to it
74. Operation of ICON
ICON
End-users
Replicate
Discover
Edge/Fog
I. Initially, ICON is in the cloud
One or multiple origination points
II. ICON monitors incoming flows
Where user requests are coming from?
III. ICON discovers deployment locations
In the domain of end-users or on a path to it
IV. ICON can take autonomous decisions
i. Deploy replica of itself
75. Operation of ICON
ICON
End-users
Use Edge
Discover
Edge/Fog
I. Initially, ICON is in the cloud
One or multiple origination points
II. ICON monitors incoming flows
Where user requests are coming from?
III. ICON discovers deployment locations
In the domain of end-users or on a path to it
IV. ICON can take autonomous decisions
i. Deploy replica of itself
76. I. Initially, ICON is in the cloud
One or multiple origination points
II. ICON monitors incoming flows
Where user requests are coming from?
III. ICON discovers deployment locations
In the domain of end-users or on a path to it
IV. ICON can take autonomous decisions
i. Deploy replica of itself
ii. Migrate closer to the end-users
ICON
End-users
Migrate
Operation of ICON
Edge/Fog
77. Salient Features of ICONs
• Capability to automatically move application services across
administrative boundaries
• Self-organizing architecture with zero management overhead
• Developers can tune the Quality-of-Experience by easily setting
budget and latency weights
• Hierarchical overlay with minimal communication overhead
• Automatic termination with budget reallocation based on thresholds