1. SoftMoW: Recursive and Reconfigurable Cellular WAN Architecture
Mehrdad Moradi Wenfei Wu Erran Li Z. Morley Mao

2. Current Mobile WANs Organized into rigid and very large regions
Minimal interactions among regions
Centralized policy enforcement at PGWs
UE1
UE2
UE3
Two Regions
UE4

3. Mobile WAN Problems Suboptimal routing in large carriers
Lack of sufficiently close PGW (Path Inflation, PAM’14)
Lack of support for seamless inter-region mobility
No inter-PGW mobility support (DMM, Zuniga et.al., 2013)
Scalability and reliability
Centralized policy enforcement
Ill-suited to adapt to new trends of mobile traffic
Signaling storm problem
is a major cause of path inflation

4. What is SoftMoW? Clean-slate architecture of cellular WANs
Scalable control plane and data plane
Performs new global applications
Runs Region optimization
Supports Seamless mobility
Enables optimal end to end paths
Goal: millions of UEs and hundreds of thousands of Base stations

5. SoftMoW’s Cellular WAN
new global cellular apps
avoid path inflation
Seamlessly interconnected
SDN switches nationwide
Decentralized policy enforcement
Sufficient # egress points
middle-boxes placed
in edge networks
Base stations organized into groups
Fine-grained classifier per BS
(e.g., SofCell)

6. Challenge 1:Scalability
O(100K) data plane switches, base stations, and middleboxes
O(100M) UEs
distributed over a continent
Scalable control plane?
multiple controllers organized in flat topologies
SoftMoW
HyperFlow
Hierarchical
Recursive
Reconfigurable
ONOS
Onix

7. Recursive and Reconfigurable Control Plane
Recursively, each controller in the hierarchy exposes:
Non-leaf controller can reconfigure logical data plane devices
Optimize hierarchy without global states
G-Middlebox
G-BS
G-switch
Hierarchical
Control Plane
Recursively partition the data plane network into logical regions and assign to controllers

8. Recursive Construction
Level 3 C5
G-BS3
GS 5
GS 6
Optimality? C3 C4
Level 2
Virtual Fabric: performance
metrics per G-switch port pair
GS 1
G-BS1
GS 2
G-BS2
GS 3 C0 C1 C2
Level 1
BS Group 1
SW1
BS Group 2
SW2
SW3
SW4
SW5
SW6

9. SoftMoW Controller Architecture
Network operating system (NOS)
Contains core services
Agnostic of cellular-specific applications
Operator apps
E.g., region optimization, HSS, PCRF
Recursive abstraction app (RecA)
Eastbound API for operator apps
Agent exposes G-switch, G-BS, G-Middleboxes
Management Plane
Bootstraps the recursive control plane.
E.g., IP assignment, tree configuration
Agent communicates with the parent controller

10. Challenge 2: Topology Discovery
SW1
SW2 C1 C5
Level 3
GS 5
GS 6
G-BS3
LLDP protocol C3 C4
Scalable topology discovery without breaking abstractions?
GS 1
GS 2
GS 3
Level 2
G-BS1
G-BS2
Each physical inter-switch link is visible to only one controller C0 C1 C2
SW1
SW5
SW2
SW3
SW4
Level 1
SW6
BS Group 1
BS Group 2

11. Core Service: Topology Discovery
Scalable and recursive link and switch discovery protocols
Similar to LLDP
Parallel-sequential periodical protocol
G-switch discovery
Inter-G-switch link discovery
Abstract G-switch computation

12. Core Service: Topology Discovery
In SoftMoW, each link discovery message has:
Stack field: stores the traversed path in the control plane
Stack entry format: (Controller ID, G-switch ID, G-switch port)
(C0, GS1, p1)
From (GS1,p1) to (GS2, p4) p1 C0 p4
GS1
GS2
Origination
(C1, SW2, p2)
(C0, GS1, p1) C1 C2
(GS2, p4) p2 p3
Return
SW1
SW2
(SW3, p3)
SW3
SW4
Payload
Stack

13. Challenge 3:Path Implementation
Scalable path implementation without
breaking abstractions?
Level 3
GS 5
GS 6
G-BS3
Controllers make local decisions C3 C4
Decisions made by a controller must be visible across its links
GS 1
GS 2
GS 3
Level 2
G-BS1
G-BS2 C0 C1 C2
Label stacking: packing all local decisions into each packet
SW1
SW2
SW3
SW4
SW5
Level 1
L1, L2, L3
SW6
The controller expects these two G-switches to process the traffic based on its local decisions.
However, each of these G-switches represent a controller sub-hierarchy where child controllers
Make local decisions. One can achieve this goal, by packing all decisions into each packet.
BS Group 1
BS Group 2
Per packet stack

14. Recursive Label Swapping
Fine-grained classifications at base stations
Any controller has its own local policy or label:
Ingress switch: Pop parent label, Push local labels
Egress switch: Pop local labels, Push parent label
Root Level
Parent Level
Leaf Level

15. Challenge 4: Region Optimization
Region optimization and without global states?
Level 3
GS 5
GS 6
G-BS3
Inter region handovers increase ‘’east-west’’ control plane load C3 C4
GS 1
GS 2
GS 3
Level 2
Require the intervention of at least three controllers
G-BS1
G-BS2 C0 C1 C2
Regions should be refined to reduce such loads
SW1
SW2
SW3
SW4
SW5
Level 1
SW6
Group 1
Group 2

16. App: Region Optimization and Reconfiguration
Handover patterns vary across time-of-day.
Difficult to find static borders
Design a greedy-iterative approach
Regions at a higher level has higher priority
Handover graphs to record patterns
Nodes are G-BSes
W(edge): handover frequency
Handover-specific Reconfiguration
Detach a G-BS from a source G-switch
Associate with a destination G-switch
Reconfiguration

17. App: Region Optimization and Reconfiguration
GS 2
G-BS2
SW1
GS 1
G-BS1
SW2
SW3
Group 1
Group 2
Initiator controller finds the G-BS with the highest gain
Gain: Inter-region handover load reduction
Contact the management plane (MP)
MP Finds the leaf controllers
Seamless control transfer at the leaf
EQUAL ROLE
Logical regions updated bottom-up
Stop when:
There is no gain
All G-switches (recursive child regions) meet their max and min loads C3
G-BS1
GS 1
GS 2
G-BS2 C0 C1
SW1
SW2
SW3
Group 1
Group 2

18. Evaluation Prototype SoftMoW on top of the Floodlight and Mininet
Egress points using iPlane traces
Core topology using RocketFuel
Two-level SoftMoW
Large scale simulation based on data from a LTE network
~1000 base stations and 1 million mobile devices
Inferred BS groups based on handover patterns
Generate uplink/downlink traffic based on actual demands
Attach base station groups to access switches

19. Recursive Topology Discovery
A 2-level SoftMoW compared with flat topology discovery
SoftMoW’s controllers discover between 44% and 58% faster
Root cause: less queuing delay at each due to information hiding

20. Performance A 2-level SoftMoW compared with LTE network
Global routing for variable number of egress points
End-to-end shortest paths
75th and 85th percentile RTT latencies reduce by 43% and 60%
End-to-End Hop Count
End-to-End Latency

21. Reconfiguration and Region Optimization
Inter-region handovers handled by the root over 48 hours
8 leaf region and 4 leaf region settings.
High inter-region handovers in peak hours (e.g., inter-city movements)
Root can reduce the load by 38.08% to 44.61%

22. Conclusion
SoftMoW:
A scalable architecture that is based on on effective recursive and reconfigurable abstractions
Three core services: link discovery, routing, and path setup
Three operator apps: global mobility, region optimization, and bearer management