adaptive source routing and speculative execution … · bachelor thesis oliver michel...

ADAPTIVE SOURCE ROUTING AND SPECULATIVE EXECUTION FOR MULTI-HOMED WIRELESS CLIENTS IN PRECLINICAL MEDICAL CARE

Bachelor Thesis

Oliver [email protected]

November 2012

Monday, February 11, 13

mailto:[email protected]

mailto:[email protected]


Outline

Oliver Michel, University of Vienna, 11/30/2012

•Motivation

•Objectives and Measures

•Prototype

•Evaluation

•Conclusions


Outline

Oliver Michel, University of Vienna, 11/30/2012 2Monday, February 11, 13


Motivation 1


• tele-medicine becoming more popular in all medical disciplines and application scenarios

3

[Bergrath et.al. - PLOS ONE 2012]

Monday, February 11, 13

•not widely deployed in pre-hospital medical care due to deficient and constrained mobile connectivity

•design an easy to deploy framework usable for any application to enhance mobile network connectivity

• special needs in health-care scenario


Motivation 2



Objective


What to achieve?

5Monday, February 11, 13

•always available

• responsive

• fast

• secure


Objective


What to achieve?


•always available

• responsive

• fast

• secure


Objective


• resilience

•availability

• latency

• throughput

• transparency

What to achieve?


•always available

• responsive

• fast

• secure


Objective


• resilience

•availability

• latency

• throughput

• transparency

•multi-homing

• resource selectionsource-routing

•concurrent multipath transmission

• speculative execution

What to achieve?



Multi-Homing



Multi-Homing


ISP 1 ISP 2 ISP 3



Multi-Homing


ISP 1 ISP 2 ISP 3ISP 1 ISP 2 ISP 3



Multi-Homing


ISP 1 ISP 2 ISP 3ISP 1 ISP 2 ISP 3ISP 1 ISP 2 ISP 3



Multi-Homing


ISP 1 ISP 2 ISP 3ISP 1 ISP 2 ISP 3ISP 1 ISP 2 ISP 3ISP 1 ISP 2 ISP 3



Resource-Selection


ISP 1

ISP 2

ISP 3



Concurrent Multipath Transmission


ISP 1

ISP 2

ISP 3



Speculative Execution


ISP 1

ISP 2

ISP 3

3

32

21

1



Latency-based Path Selection


Lr =

0

BBBB@

��!lr�!...

��!lr�1�!lr

1

CCCCA�!lr = (lr,i, lr,i+1, ..., lr,n)

s(L, µ) = (p0, p1, ..., pµ), µ n

min(��!lr+1) = lr+1,sr(L,µ)

path latencies get collected in sliding window matrix:

selection function s(L, μ) returns “best” μ path indices according to some objective (here min! latency):

path decision was correct if:



Requirements


• improve data-transmission between a mobileclient and arbitrary servers

• transparently usable by any application

•executable in user-space



System Architecture


mobile clientwith path-selection agent

ISP 1

ISP 2

ISP 3servergateway



System Architecture: Client


Client

wlan0192.168.60.2/24

wlan1192.168.61.2/24

wlan2192.168.62.2/24

Clienttun010.0.0.1/24

A1

A2

A3

•controls active probing

•performs path-selection

• reads data from tunnel-device

•encapsulates data and sends over selected path


• reads data from public interface

• removes custom header

•writes data to tunnel device for forwarding in kernel and NAT

• (also does active probing on return path)


System Architecture: Gateway


Gateway

GW tun010.0.0.2/24 FW NAT eth0

192.168.1.5/24eth1

192.168.88.2/24



System Architecture: Packet Encapsulation


ASRTunnelClient ASRTunnelGateway NAT

CPi ➞ GPCT ➞ SCT ➞ S CT ➞ S GP➞ S

GP ➞ CPi S ➞ CTS ➞ CT S ➞ CT S ➞ GP

TunnelDeviceCT

Public InterfaceS

• IP-layer packets get encapsulated as UDP payload for transmission over tunnel

•custom 12 Byte packet header: uint8_t _flags; uint8_t _client_id; uint16_t _pl_length; uint32_t _seqn; uint32_t _client_addr;



System Architecture: Modules


NAT

Public InterfaceGP

DatagramSocketGP

ASRTunnelGateway

TunnelDeviceGT

Public InterfaceSP

Server Application

ASRTunnelClient

PathSelector

SMAPathSelector

TunnelDeviceCT

DatagramSocketCPi

Client

Gateway

Server

Agent

PathSelector

SMAPathSelector

AgentInternet

Internet

Client ApplicationClient

ApplicationClient Application

RoutingRouting

Routing

RoutingRouting

IP Rules

RoutingRouting

IP Tables



Results: Path Selection 1


2 4 6 8 10 12 140.0

0.1

0.2

0.3

0.4

0.5

window size

fractionofpathchanges

2 4 6 8 10 12 140.70

0.75

0.80

0.85

0.90

0.95

1.00

window size

fractionofcorrectdecisions

2 4 6 8 10 12 14

80

90

100

110

window size

latency@msD 99th %ile

95th %ile90th %ilemean

• post-hoc analysis of SMA, TPMA, and EWMA algorithms on measured latency data (PlanetLab) and sampled values from Pareto-Type 1 and Gaussian

• for window-based algorithms a window size between 4-6 seems reasonable



Results: Path Selection 2


100 150 200 250 300 3500.0010.002

0.0050.010.02

0.050.10.2

0.51

Latency @msD

FractionofpacketslateHlog

.L0 10000 20000 30000 40000 50000

0.0010.002

0.0050.010.02

0.050.10.2

0.51

Latency @msD

FractionofpacketslateHlog

.L

• SMA5 path-selection from data collected from a PL tripartide overlay graph

• SMA surprisingly good, close to optimal (in retrospect)

• TPMA, EWMA slightly inferior

• SMA5 using data-flow (64-128kbit/s) on wireless testbed

• latencies of probe, as well as data traffic (1 per second, exponential)



Results: Fast Adaption and Handover


0 100 200 300 4000

1000

2000

3000

4000

5000

6000

7000

Time @sD

Throughput@By

teD

0 20 40 60 80 100 120 1400

20000

40000

60000

80000

100000

Packet

Latency@msD

•measured latency data modified using netem

• satisfying handover-behavior

• little occurrence of packet losses while switching paths

• continuous high data throughput



Results: Packet Replication


1e-05

0.0001

0.001

0.01

0.1

1

0 200 400 600 800 1000fr

acti

on o

f pac

kets

late

latency threshold [ms]

1 path2 paths3 paths

• redundant multipath transmission (i.e. packet duplication) may significantly reduce the latency tail

• virtually no packet loss in case of path failure

• general technique for system designers when secondary resources would be idling normally [Vulimiri et. al. - Hotnets ’12]



Future Directions


• detailed Investigation of TCP behavior

- TCP connections sometimes get stuck when paths switch frequently

• IP Fragmentation

- iptables NAT functionality does not support fragmented IP packets

• more detailed parameter study for path-selection

- other algorithms (e.g. multi-armed bandit)

• experimentation with different probing intervals

• consideration of further metrics (esp. bandwidth)



Conclusions


• transparently usable framework designed to enhance mobile data connectivity

• technically feasible and performant

•modular design allowing system designers to attach custom logic and state

• investigation of reasonable approximations for the Internet path-selection problem

• focus on special needs and applicability in the health-care field (esp. preclinical medical care)


adaptive source routing and speculative execution … · bachelor thesis oliver michel...

Documents