I. Matta 1

On the Cost of Supporting Mobility and Multihoming

Vatche Ishakian, Ibrahim Matta, Joseph AkinwumiComputer ScienceBoston University

I. Matta

Mobility = Dynamic Multihoming Hosts / ASes became increasingly multihomed Multihoming is a special case of mobility

RINA (Recursive InterNetwork Architecture) is a clean-slate design – http://csr.bu.edu/rina

RINA routing is based on node addresses Late binding of node address to point-of-attachment

Compare to LISP (early binding) and Mobile-IP Average-case communication cost analysis Simulation over Internet-like topologies

What’s wrong today?one big, flat, open net

Network

Transport

Data Link

Physical

Applications

Network

Transport

Data Link

Physical

Applications

Network

DL DL

PHY PHY

Web, email, ftp, …

There’s no building block We named and addressed the wrong things (i.e. interfaces) We exposed addresses to applications

www.cs.bu.edu128.197.15.10

128.197.15.

1

128.

10.1

27.2

5

128.10.0.0 128.197.0.0

TCP, UDP, … IP protocol

RINA offers better scoping

Network

Transport

Data Link

Physical

Applications

Network

Transport

Data Link

Physical

Applications

Network

DL DL

PHY PHY

TCP, UDP, …

IP

Web, email, ftp, …

IPC Layer

IPC Layer IPC Layer

The IPC Layer is the building block and can be composed An IPC Layer has all what is needed to manage a “private” network,

i.e. it integrates routing, transport and management E2E (end-to-end principle) is not relevant

Each IPC Layer provides (transport) service / QoS over its scope IPv6 is/was a waste of time!

We can have many layers without too many addresses per layer

5

RINA: Good Addressing – private mgmt

Destination application is identified by “name” Each IPC Layer is privately managed

It assigns private node addresses to IPC processes It internally maps app/service name to node address

BA

I1I2

want to send message to “Bob”

IPC Layer

To: B

“Bob”B

Bob

IPC Layer

6

RINA: Good Addressing - late binding

Addressing is relative: node address is name for lower IPC Layer, and point-of-attachment (PoA) for higher IPC Layer

Late binding of node name to a PoA address A machine subscribes to different IPC Layers

BA

I1 I2

want to send message to “Bob”

BI2

To: B

Bob

IPC Layer

IPC Layer

B, , areIPC processeson same machine

I1 I2

I. Matta 7

RINA: Good Routing

Back to naming-addressing basics [Saltzer ’82] Service name (location-independent) node name (location-dependent) PoA address (path-dependent) path

We clearly distinguish the last 2 mappings Route: sequence of node names (addresses) Late binding: map next-hop’s node name to PoA at lower IPC

level

source destination

8

Mobility is Inherent

Mobile joins new IPC Layers and leaves old ones Local movement results in local routing updates

CHMH

9

Mobility is Inherent

Mobile joins new IPC Layers and leaves old ones Local movement results in local routing updates

CH

10

Mobility is Inherent

Mobile joins new IPC Layers and leaves old ones Local movement results in local routing updates

CH

I. Matta

Compare to loc/id split (1) Basis of solutions to the multihoming issue Claim: the IP address semantics are overloaded as both

location and identifier LISP (Location ID Separation Protocol) ’06

EIDx EIDy

EIDx -> EIDy

EIDx EIDy

RLOC1x RLOC2y

Mapping: EIDy RLOC2y

11

Compare to loc/id split (2) Ingress Border Router maps ID to loc, which is the

location of destination Egress BR Problem: loc is path-dependent, does not name the

ultimate destination EIDx -> EIDy

EIDx EIDy

RLOC1x RLOC2y

Mapping: EIDy RLOC2y

12

I. Matta

LISP vs. RINA vs. … Total Cost per loc / interface change =

Cost of Loc / Routing Update +

[ Pcons*DeliveryCost + (1-Pcons)*InconsistencyCost ]

: expected packets per loc change

Pcons: probability of no loc change since last pkt delivery

RINA’s routing modeled over a binary tree of IPC Layers: update at top level involves route propagation over the whole network diameter D; update at leaf involves route propagation over D/2h, h is tree height

13

I. Matta

LISP

14

I. Matta

LISP

15

I. Matta

RINA

16

I. Matta

RINA

17

I. Matta

RINA

18

I. Matta

MobileIP

19

I. Matta

LISP vs. RINA vs. …

RINA

8x8 Grid TopologyRINA uses 5 IPC levels; on average, 3 levels get affected per move

LISP

20

Simulation: Packet Delivery Ratio

BRITE generated 2-level topology

Average path length 14 hops

Random walk mobility model

Download BRITE from www.cs.bu.edu/brite

I. Matta 21

RINA

LISP

Simulation: Packet Delay

I. Matta 22

LISP

RINA

I. Matta

Bottom Line: RINA is less costly

RINA inherently limits the scope of location update & inconsistency

RINA uses “direct” routing to destination node

More work: prototyping

23

I. Matta

RINA papers @http://csr.bu.edu/rina

Thank You

Questions?

24