Network Coding Overview NWRC meeting March 11, 2013

Size: px

Start display at page:

Download "Network Coding Overview NWRC meeting March 11, 2013"

Herbert Barnett
5 years ago
Views:

1 Network Coding Overview NWRC meeting March 11, 2013

2 Coding à Network Coding Coding Today (all end- to- end) Coding Tomorrow (using Random Linear Network Coding) Network Classical Mul<path Mul<cast Mul<source Mul<- des<na<on / Mesh 3

3 Why RNLC is Different Code Capabili<es Rateless (e.g. Fountain) Codes Block Codes Erasure correc<on Code is carried within each packet Completely distributed opera<on De- code using unencoded packets Able to generate valid codes from coded or unencoded packets Composability without decoding (add incremental redundancy) Encode data in a sliding window 4

4 Ini<al Stack Implementa<ons Applica<on Layer TCP Socket Opera<ng System Agnos<c Implementa<on (So$ware) TCP UDP IP Layer Network Layer Physical Layer Kernel Implementa<on (So$ware) Chip Implementa<on (Hardware) Implementable at any layer as a patch 5

5 Applica<on Layer TCP Case Study Classical Network Applica<on Layer TCP Socket Mul0path TCP UDP Mul0source Mul0- des0na0on IP Layer Network Layer Physical Layer 6

6 TCP Ill- Suited to Random Losses sender window TCP p 1 p 2 p 3 p p 1 4 p 2 ACK(p 1 ) p 3 p 4 ACK(p 1 ) ACK(p 1 ) p 2 p 3 p 4 p 5 p 5 ACK(p 1 ) p 2 p 3 Window closing: W W/2 Triple- duplicate ACKs! Can t increment window by 1 Par<al sliding window 7

7 Coded TCP Maintains Throughput sender window p 1 p 2 p 3 p 4 p 4 p 5 p 6 p 7 p 7 p 8 p 9 p 10 p 11 p 8 p 9 p 10 p 11 p 12 Σp Σp Σp Σp Σp Σp Σp Σp Coded TCP SEEN(1) SEEN(2) SEEN(3) SEEN(4) SEEN(5) SEEN(6) SEEN(7) Prevents random losses being interpreted as conges<on! There is a lag in the SEEN acks: To avoid lag, introduce redundancy! 8 Minij Kim, Muriel Médard, João Barros Modeling network coded TCP throughput: a simple model and its valida<on Valuetools, (2010)

8 Performance With Conges<on Losses sender window p 1 TCP sender window p 1 p 2 p 3 p 4 p 1 p 2 p 3 p p 4 2 ACK(p 1 ) p 3 p 4 Σp Σp Σp Σp TCP/NC SEEN(1) p 2 p 3 p 4 p 5 Wai<ng p 2 p 3 p 4 p 5 Wai<ng p 2 Window closing: W 1 Time- out! p 2 Window closing: W 1 Time- out! 9 S<ll allows conges<on control while masking random losses!

9 Experimental Results: Single Path Server: Amazon EC2 instance in CA Client: Desktop at RLE MIT, using WiFi(s) > 100 ms wlan0 Loss rate 0% 1% 2% 3% 4% 5% TCP (Mbps) CTCP (Mbps) Avg. Throughput (Mbps) Coded TCP TCP 0% 1% 2% 3% 4% 5% Loss Rate 10 6 seconds to transmit 130 seconds to transmit Leo Urbina Applying Network Coding to TCP MIT M.Eng. Thesis 2012

10 Experimental Results: Single Path Average Throughput: 0.77 Mbps Average Throughput: Mbps Example: 2% loss 11

11 Testbed Measurements 60 s video Full download 60 s video Progressive download 12 Minji Kim, Jason Cloud, Ali ParandehGheibi, Leonardo Urbina, Kerim Fouli, Douglas Leith, Muriel Médard Network Coded TCP (CTCP) arxiv: v2

12 WiMax Experiments 13

13 Coded TCP Network Improvements Classical Network Applica<on Layer TCP Socket Mul0path TCP UDP Mul0source Mul0- des0na0on IP Layer Network Layer Physical Layer 14

TCP) End- to- End Coding (including Fountain Codes) 0.9 0.

14 Coded TCP (CTCP) Over a Network source 10% loss 10% loss 10% loss 10% loss client Longer paths lead to higher losses, resul<ng in poor performance Best Possible Throughput Rate (using Coded TCP) End- to- End Coding (including Fountain Codes) TCP s performance degrades super- linearly with end- to- end loss rate à benefits of Coded TCP are even greater in this scenario 15

Randomized Composability: Distributed Redundancy Network When Coding Needed P 1+ P2+4 P3+ 2 P4 20% Loss Node 2 P 1 + 8 P 2 + 9 P 3 +P 2 8 P 1 + 6 P 2 + P 3 + P 4

packets) 3P 1+ 5P 2+ 2P 3+ 7P 4 P 1+ P2+ 4 P 3+ 2 P4 2 P 1 + 8 P 2 + 9 P 3 +P 2 8 P 1 + 6 P 2 + P 3 + P 4 7 P 1 +2 P 2 +3 P 3 +4 P 4 2 P 1 + 8 P 2 + 9 P 3 +P 2

15 Randomized Composability: Distributed Redundancy Network When Coding Needed P 1+ P2+4 P3+ 2 P4 20% Loss Node 2 P P P 3 +P 2 8 P P 2 + P 3 + P 4 7 P 1 +2 P 2 +3 P 3 +4 P 4 Over reliable network send minimum packets (4 variables, 4 packets) Re- encode à add a packet without de- coding (4 variables, 5 packets) 3P 1+ 5P 2+ 2P 3+ 7P 4 P 1+ P2+ 4 P 3+ 2 P4 2 P P P 3 +P 2 8 P P 2 + P 3 + P 4 7 P 1 +2 P 2 +3 P 3 +4 P 4 2 P P P 3 +P 2 3P 1+ 5P 2+ 2P 3+ 7P 4 8 P P 2 + P 3 + P 4 7 P 1 +2 P 2 +3 P 3 +4 P 4 Needed packets arrive (4 variables, 4 packets) allows optimized packet sends at any node 16

16 Coding Coefficients Carried within Packet 17 Jay Kumar Sundararajan, Devavrat Shah, Muriel Médard, Szymon Jakubczak, Michael Mitzenmacher, João Barros, Network Coding Meets TCP: Theory and Implementa<on. Proceedings of the IEEE 99 (3): (2011)

17 Mul<hop Wireless Network FTP sender FTP receiver From given data 18

18 Performance Comparison TCP End-to-end coding Re-encoding at node 3 only Mbps Mbps Mbps Time average throughput (over 641 seconds) (assuming each link has a bandwidth of 1 Mbps in the absence of erasures) 19

19 CTCP versus Mul<path TCP Using Rou<ng 10% loss on each link - 10% - 10% - 10% - 10% - 10% source - 10% - 10% - 10% - 10% client - 10% - 10% - 10% - 10% MPTCP may poten<ally provide mul<- path communica<on BUT: Difficult and complex scheduling at the source needed Round- robin scheduling is inefficient CTCP provides mul<- path communica<on throughput of 2.7 without complex scheduling at the source 20

20 Experimental Results: Mul<ple Path Server: Amazon EC2 instance in CA Client: Desktop at RLE MIT, using WiFi(s) Limited each path to < 8-9 Mbps > 100 ms wlan0 wlan1 Loss rate 0% 1% 2% 3% 4% 5% wlan0 (Mbps) wlan1 (Mbps) CTCP (Mbps) Example: 0% loss Combined throughput at Client Throughput on wlan0 21

21 Applica<on Layer Base Sta<on Case Study Classical Network Applica<on Layer TCP Socket Mul0path TCP UDP Mul0source Mul0- des0na0on IP Layer Network Layer Physical Layer 22

Value of Goodput with User Growth Problem Statement: As # of users grows, addi<onal base sta<ons are added for throughput, not coverage MinJi Kim Thierry Klein Emina Soljanin João Barros Muriel

22 Value of Goodput with User Growth Problem Statement: As # of users grows, addi<onal base sta<ons are added for throughput, not coverage MinJi Kim Thierry Klein Emina Soljanin João Barros Muriel Médard Modeling Network Coded TCP: Analysis of Throughput and Energy Cost hqp://arxiv.org/abs/ CoRR abs/ db/journals/corr/corr1208.html#abs improved goodput = fewer Base Stations 23

23 Expands Base Sta<on Reach Joint research with MIT and Alcatel- Lucent 5.08 MB # of Base Sta<ons Increasing User Ac<vity With operators lower both OpEx and CapEx 24

24 Link Layer Implementa<on Case Study Classical Network Applica<on Layer TCP Socket Mul0path TCP UDP Mul0source Mul0- des0na0on IP Layer Network Layer Physical Layer 25

25 Network Coding vs. ARQ and HARQ Scenario: 5- packet block transfer from BS to SS Downlink: fixed 40% packet error paqern (every 3 rd and 5 th packet) Uplink: feedback NACKs not subject to loss ARQ: repeated transmissions create RTT feedback loops HARQ: feedback reduced by combining corrupted packet versions Network Coding: a- priori systema<c coding with added redundancy of 3/5 Clear delay, throughput, and energy gains No feedback loop Redundancy cost amor<zed over block 26 Surat Teerapiqayanon, Kerim Fouli, Muriel Médard, Marie- José Montpe<t, Xiaomeng Shi, Ivan Seskar, Abhimanyu Gosain Network Coding as a WIMAX Link Reliability Mechanism MACOM 2012L 1-12 Surat Teerapiqayanon, Kerim Fouli, Muriel Médard, Marie- José Montpe<t, Xiaomeng Shi, Ivan Seskar, Abhimanyu Gosain Network Coding as a WIMAX Link Reliability Mechanism: An Experimental Demonstra<on MACOM 2012: 75-78

26 Experimental Setup Intra- flow NC modules at the Base Sta<on (BS) and Subscriber Sta<on (SS) Toggle ARQ, HARQ, and various NC configura<ons IPERF à applica<on- layer throughput / loss UFTP (FTP over UDP) à applica<on- layer file- transfer delay 27

IP- Based Implementa<on Remote access of enode B Card driver WiMAX MAC inaccessible à IP- based implementa<on Performance measurements at the applica<on layer (IPERF and UFTP) IP layer: Ne=ilter used

27 IP- Based Implementa<on Remote access of enode B Card driver WiMAX MAC inaccessible à IP- based implementa<on Performance measurements at the applica<on layer (IPERF and UFTP) IP layer: Ne=ilter used to intercept packets, route them to encoder/decoder, then re- inject them to IP layer PDCP does not need to be involved, although that may be quite doable Network coding included at the e- Node B before handing it to the MAC, and ARQ and HARQ bypassed at the MAC does not require a proxy, but can be used if convenient. Occurs below transport, so not touching TCP per se 28

28 Consistency NC- Best decreases packet loss from 11-32% to nearly 0% NC offers up to 5.9x gain in throughput and 5.5x reduc<on in file transfer delay Loss Throughput File transfer delay 29

29 Looking Forward The applica<ons of network coding are varied and changing This is indeed the right <me for a standardiza<on effort Network coding can be implemented at different layers and in different por<ons of the network Standardiza<on should allow the flexibility needed to allow these diverse implementa<ons 30

Network Coding and Reliable Communications Group 1. Collaborators

Network Coding and Reliable Communications Group 1. Collaborators 1 Collaborators MIT: Jason Cloud, Flavio du Pin Calmon, Szymon Jakubczak (now Google), Minji Kim (now Akamai), Marie- Jose Montpetit, Asuman Ozdaglar, Ali ParandehGheibi,, Devavrat Shah, Jay-Kumar Sundararajan