1. Redesigning MPTCP for Edge
Clouds
ACM Mobicom 2018 Student Research Competition
Delhi, India
Nitinder Mohan
⊥
Tanya Shreedhar
∤
Aleksandr Zavodovski
⊥
Otto Waltari
⊥
Jussi Kangasharju
⊥
Sanjit K. Kaul
∤
⊥
University of Helsinki, Finland
∤
IIIT Delhi, India
2. Edge Networks
• Small-scale server(s) equipped with multiple NIC at network “edge”
• Distribute data with other edge servers over last-mile wireless
links
• Support for cloud technologies and scenarios such as
container/VM migration etc. developed for managed networks
Edge-to-Edge networks must ensure network reliability from
A1
A2
Core
Network
B2
B1
VM
2
3. • Standardized kernel extension to TCP
• Forms multiple TCP flows over all available network interfaces
to simultaneously utilize them
Improvements: Increased robustness, bandwidth aggregation,
seamless handovers etc.
Multipath-TCP on Edge
MPTCP
TCP
Flow 1
A1
A2
Core
Network
B2
B1
TCP
Flow 1
TCP
Flow 2
TCP
Flow 2
VM
MPTCP
SRTTflow1
SRTTflow2
?SRTTflow1
<
SRTTflow2?
3
4. Default MPTCP Performance
How does MPTCP perform in real Edge ↔ Edge
scenario?
Setup: Both Sender and Receiver are equipped with two 802.11g WiFi over
MPTCP v0.94
Network test: B2 interface at receiver starts to experience packet loss with
20% probability
MPTCP
TCP
Flow 1
A1
A2
Core
Network
B2
B1
TCP
Flow 1
TCP
Flow 2
TCP
Flow 2
MPTCP
4
7. MPTCP Sub-optimal Flow Decisions
1. SRTT is a delayed metric for estimating network conditions
• Cannot pin-point exact source of delay in network
• Cannot react to cause of delay
2. Little to no interaction between individual flows
• Highly dependent on individual TCP flow throttling and adjustment
3. Sender is King!
Current MPTCP is a dumb collection of multiple
individual TCP connections
7
8. Receiver-Assisted MPTCP (RAMPTCP)
“Reliability-first” extension to MPTCP by reducing
packet re-transmission and re-ordering
1. Enables receiver to send its readily-available last-mile
characteristics to sender in ACK
e.g. channel utilization, signal strength, path loss% etc.
2. Sender can use extra information along with SRTT to better
estimate network conditions
8
9. RAMPTCP in Action
Data Delivery
Reverse-Path Information
Delivery
Per-Segment Delay
Estimation
Flow Control Actions
9
10. Path 2
A2
A1 B1
B2
Path 1
Data Delivery
Reverse-Path Information
Delivery
Per-Segment Delay
EstimationFlow Control Actions
Seq: 1
Data
Seq: 2
Data
10
11. Path 2
A2
A1 B1
B2
Path 1
Seq: 1
B1:
B2:
Data Delivery
Reverse-Path Information
Delivery
Per-Segment Delay
EstimationFlow Control Actions
10
12. Path 2
A2
A1 B1
B2
Path 1
RTTpath = Tsender + Tcore + Treceiver
Data Delivery
Reverse-Path Information
Delivery
Per-Segment Delay
EstimationFlow Control Actions
10
13. Path 2
A2
A1 B1
B2
Path 1
Data Delivery
Reverse-Path Information
Delivery
Per-Segment Delay
EstimationFlow Control Actions
Seq: 3
DataSeq: 4
Data
Possible control actions:
Limit packets on flow, Out-of-order injection,
Packet duplication, Change TCP send rate,
Boycott flow usage, …
10
14. Preliminary Evaluation
ns3 Simulation based on Direct Code Emulation (DCE)
B1 interface suffers packet losses due to interference from 2-5sMPTCP
TCP
Flow 1
A1
A2
Core
Network
B2
B1
TCP
Flow 1
TCP
Flow 2
TCP
Flow 2
MPTCP
11
0
5
10
15
Throughput(Mbps)
0 1 2 3 4 5 6 7 8 9 10
Time (secs)
0
5
10
15
Flow 1
Flow 2
MPTCP
RAMPTCP
≈19% increase in
throughput
≈ 58% reductions in
retrasmissions
15. Discussion and Future Work
1. Incorporate any access-layer technology i.e. WiFi, Ethernet,
LTE, 5G etc.
2. How to effectively send receiver-side information to sender?
3. Effective RAMPTCP control decisions?
RaspberryPi
Battery
LTE modem 12
17. Rise of Edge Clouds
Network
DatacenterEdge
Server
User
→ Small servers deployed at network
edge to compute user data
→ Equipped with multiple NIC with
different access technology e.g. WiFi,
LTE, Ethernet etc.
Motivation:
Decreased latency and network traffic
Computing data of local relevance
Low network delay requirements 17
18. Preliminary Evaluation
ns3 Simulation based on Direct Code Emulation (DCE)
1. Both sender and receiver are running MPTCP v0.89
2. B1 interface suffers packet losses due to interference from 2-
5s
RAMPTCP configuration:
→ Sender is aware of receiver channel utilization, SNR, path loss
%
MPTCP
TCP
Flow 1
A1
A2
Core
Network
B2
B1
TCP
Flow 1
TCP
Flow 2
TCP
Flow 2
MPTCP
11
First we talk about edge servers and overview of what researchers have proposed
Design criteria:
Follows principle of TCP congestion control design
Metric must be readily extracted from Linux kernel
aims to improve network reliability irrespective of use-case
First point
Identify properties and behavior of differentaccess technologies
Extract equivalent path characteristics from Linux Kernel
second point
Available space in TCP packet?
Available space in MPTCP Data Sequence Signal (DSS) packet?
New packet?
third point
Handle intermittent delays
Perform better than single-path TCP
Avoid complete loss of service
First we talk about edge servers and overview of what researchers have proposed