internet design case study: ip multicast

Spring 2007 CSci5221: IP Multicast 1

Internet Design Case Study:IP Multicast

• What is multicast? Why multicast?• Understand the “spirit”, not necessarily the detail• Some design/implementation principles we’ll use:

– indirection– soft state– randomization– …

• Evolution of IP Multicast and Lessons learned


Motivation• Often want to send data to many machines at once

– Video distribution (TV broadcast)– Teleconferences, etc.– News updates

• Using unicast to reach each individual is hard and wasteful– Sender state: ~O(n) and highly dynamic– Total load: ~O(nd) where d is net diameter– Hotspot load: load ~O(n) on host and first link

• Multicast:– Sender state: O(1), total load O(d log n), hotspot load O(1)


Multicast: one sender to many receivers

• Multicast: act of sending datagram to multiple receivers with single “transmit” operation

• Question: how to achieve multicast

Network multicast• router actively participate in

multicast, making copies of packets as needed and forwarding towards multicast receivers

Multicastrouters (red) duplicate and forward multicast datagrams


Indirection

Indirection: rather than reference an entity directly, reference it (“indirectly) via another entity, which in turn can or will access the original entity

"Every problem in computerscience can be solved by adding another level of indirection" -- Butler Lampson

A

B

x


Multicast via Indirection

multicast group concept: represented by a “logical” (multicast) address -- use of indirection– sender sends IP datagrams to the “logical” multicast

address– routers forward multicast datagrams to hosts that have

“joined” that multicast group

128.119.40.186

128.59.16.12

128.34.108.63

128.34.108.60

multicast group

226.17.30.197


Issues and Challenges• How does a host (user) join in a multicast group?• What to do to leave the group? • Who manages group membership dynamics?

– Membership is dynamic, but should it be open, i.e., anyone can join a multicast group?

• Who can send to the multicast group?• What state does a router need to maintain?

– When a router receives an IP datagram destined to a multicast group, where should it forward to?

• Scalability:– With number of members in a group; also with number of groups

• Other issues: – address allocation, mapping applications to multicast groups


Design Alternatives• Telephone/ATM Model:

– source/root-based: “one-to-many” multicast • mcast group associated and controlled by source/root

– receiver joins a group by sending request to source/root

– routing state installed during the “signaling” process– routing entry: [mcast address, {interface1,…,

interfacek}]

• IP Multicast Model– “many-to-many” multicast– open membership: anyone can join and leave at will– also anyone (not necessarily part of mcast) can send!– best-effort delivery


IP Multicast Service Model (rfc1112)

• Each group identified by a single IP address• Groups may be of any size• Members of groups may be located anywhere

in the Internet• Members of groups can join and leave at will• Senders need not be members• Group membership not known explicitly • Analogy:

– Each multicast address is like a radio frequency, on which anyone can transmit, and to which anyone can tune-in.


IP Multicast Addresses

• Class D IP addresses– 224.0.0.0 – 239.255.255.255

• How to allocated these addresses?– Well-known multicast addresses, assigned by IANA– Transient multicast addresses, assigned and

reclaimed dynamically, e.g., by “sdr” program

1 1 1 0 Group ID


IP Multicast Service Model (cont’d)

Division of Responsibilities:• Host’s responsibility to register interest with networks

– IGMP (Internet Group Management Protocol)• Network’s responsibility to deliver packets to host

– Multicast Routing Protocols• Left unspecified:

– Address assignment (random, MASC, etc.)– Application-to-group mapping (session directory, etc.)

Performance Goal:• roughly equivalent to unicast best-effort service in

terms of drops/delays– efficient tree– no complicated forwarding machinery, etc.

• low join/leave latency


IP Multicast Architecture

Hosts

Routers

Service model

Host-to-router protocol(IGMP)

Multicast routing protocols(various)


IP Multicast API and Multicast Routers

• Sending – same as before• Receiving – two new operations

– Join-IP-Multicast-Group(group-address, interface)– Leave-IP-Multicast-Group(group-address, interface)– Receive multicast packets for joined groups via normal IP-

Receive operation– Implemented using socket options

• Multicast Routers:– Learn of existence of multicast groups (through

advertisement)– Identify links with group members– Establish state to route packets

• replicate packets on appropriate interfaces• routing entry:

mcast addresssrc, incoming interface

List of outgoing interfaces


Internet Group Management Protocol

• End system to router protocol is IGMP• Each host keeps track of which mcast groups

are subscribed to– Socket API informs IGMP process of all joins

• Objective is to keep router up-to-date with group membership of entire LAN– Routers need not know who all the members are,

only that members exist


How IGMP Works

• On each link, one router is elected the “querier”• Querier periodically sends a Membership Query message

to the all-systems group (224.0.0.1), with TTL = 1• On receipt, hosts start random timers (between 0 and 10

seconds) for each multicast group to which they belong

QRouters:

Hosts:


How IGMP Works (cont.)

• When a host’s timer for group G expires, it sends a Membership Report to group G, with TTL = 1

• Other members of G hear the report and stop their timers• Routers hear all reports, and time out non-responding

groups

Q

G G G G

Routers:

Hosts:


How IGMP Works (cont.)• Note that, in normal case, only one

report message per group present is sent in response to a query

– Power of randomization + suppression

• Query interval is typically 60-90 seconds

• When a host first joins a group, it sends one or two immediate reports, instead of waiting for a query


Multicast Scope Control – Small TTLs

• TTL expanding-ring search to reach or find a nearby subset of a group

s

1

2

3


Multicast Scope Control – Large TTLs

• Administrative TTL Boundaries to keep multicast traffic within an administrative domain, e.g., for privacy or resource reasons

An administrative domain

TTL threshold set oninterfaces to these links,greater than the diameterof the admin. domain

The rest of the Internet


Routing Techniques• Basic objective – build distribution tree for

multicast packets

Three Basic Techniques

• Source-based trees: flood and prune– Separate shortest path tree for each sender – Begin by flooding traffic to entire network– Prune branches with no receivers– Examples: DVMRP, PIM-DM– Unwanted state where there are no receivers


Routing Techniques (cont’d)• Core or share-tree based protocols

– Specify “meeting place” aka core– Sources send initial packets to core– Receivers join group at core– Requires mapping between multicast group address and

“meeting place” – Examples: CBT, PIM-SM

• Link-state multicast protocols– Routers advertise groups for which they have receivers to

entire network– Compute trees on demand– Example: MOSPF– Unwanted state where there are no senders


Source-based Trees

Router

Source

Receiver

S

R

R

R

R

R

S

S


Shared Tree

RP

Router

Source

Receiver

S

S

S

R

R

R

R

R


Shared vs. Source-Based Trees

• Source-based trees– Shortest path trees – low delay, better load distribution– More state at routers (per-source state)– Efficient for in dense-area multicast

• Shared trees– Higher delay (bounded by factor of 2), traffic concentration– Choice of core affects efficiency– Per-group state at routers– Efficient for sparse-area multicast

• Which is better?


Distance-Vector Multicast Routing:Basics

• DVMRP consists of two major components:– A conventional distance-vector routing protocol (like

RIP) – A protocol for determining how to forward multicast

packets, based on the routing table

• DVMRP router forwards a packet if– The packet arrived from the link used to reach the

source of the packet (reverse path forwarding check – RPF)

– If downstream links have not pruned the tree


DVMRP (cont’d)• Developed as a sequence of protocols:

– Reverse Path Flooding (RPF)– Reverse Path Broadcast (RPB)– Truncated Reverse Path Broadcasting (TRPB)– Reverse Path Multicast (RPM)

• General Philosophy: multicast = pruned broadcast– Don’t construct new tree, merely prune old one

• Observation: – Unicast routing state tells router shortest path to S– Reversing direction sends packets from S without forming

loops


Basic Forwarding Rule• Routing state:

– To reach S, send along link L

• Flooding Rule:– If a packet from S is received along link L, forward on all

other links (reverse path forwarding check)

• This works fine for symmetric links– Ignore asymmetry today

• This works fine for point-to-point links– Can result in multiple packets sent on LANs


Example• Flooding can cause a given packet to be sent

multiple times over the same link

xx yy

zz

SS

a

b

duplicate packet


Broadcasting Extension• For each link, and each source S, define parent

and child– Parent: shortest path to S (ties broken arbitrarily)– All other routers on link are children

• Broadcasting rule: only parent forwards packet to L

• Problem fixed

• But this is still broadcast, not multicast!


Multicast = Pruned Broadcast

• Start with full broadcast (RPB)

• If leaf has no members, prune state– Send non-membership report (NMR)

• If all children of a router R prune, then router R sends NMR to parent

• New joins send graft to undo pruning


Example Topology

G G

S

G


Broadcast with Truncation

G G

S

G


Prune

G G

S

Prune (s,g)

Prune (s,g)

G


Graft (s,g)

Graft (s,g)

Graft

G G

S

G

G

Report (g)


Steady State

G G

S

G

G


Problems with Approach• Starting with broadcast means that all first

packets go everywhere

• If group has members on most networks, this is ok

• But if group is sparse, this is lots of wasted traffic

• What about a different approach:– Source-specific tree vs shared tree– Pruned broadcast vs explicitly constructed tree


Core Based Trees (CBT)• Ballardie, Francis, and Crowcroft,

– “Core Based Trees (CBT): An Architecture for Scalable Inter-Domain Multicast Routing”, SIGCOMM 93

• Similar to Deering’s Single-Spanning Tree• Unicast packet to core, but forwarded to multicast

group• Tree construction by receiver-based “grafts”

– One tree per group, only nodes on tree involved

• Reduce routing table state from O(S x G) to O(G)


Example

• Group members: M1, M2, M3• M1 sends data

root

M1

M2 M3

control (join) messagesdata


Disadvantages

• Sub-optimal delay

• Small, local groups with non-local core– Need good core selection– Optimal choice (computing topological center) is NP

complete


Multicast OSPF (MOSPF)

• Add-on to OSPF (Open Shortest-Path First,a link-state, intra-domain routing protocol)

• Multicast-capable routers flag link state routing advertisements

• Link-state packets include multicast group addresses to which local members have joined

• Routing algorithm augmented to compute shortest-path distribution tree from a source to any set of destinations


Example

Source 1

Receiver 1

Receiver 2

Z

W

Q

T


Link Failure/Topology Change

Source 1

Receiver 1

Receiver 2

Z

W

Q

T


Membership Change

Source 1

Receiver 1

Receiver 2

Z

W

Q

T

Receiver 3


Impact on Route Computation

• Can’t pre-compute multicast trees for all possible sources

• Compute on demand when first packet from a source S to a group G arrives

• New link-state advertisement– May lead to addition or deletion of outgoing

interfaces if it contains different group addresses– May lead to re-computation of entire tree if links are

changed


Discussion• IP anycast service works fine within a network • What about global anycast service (across domains)

– Available in special cases: • Anycast address for DNS servers• AS 112 project:

– In general, not scalable at a global scale– Some recently proposed solutions: GIA, PIAS

• Impact on UDP/TCP and Applications: anycast addr appear same as unicast to senders!– UDP stateless!

• Okay if applications do not maintain state over multiple datagrams; Example: DNS

• problematic otherwise: e.g., multimedia applications over UDP – TCP stateful: how to establish a connection with anycast

address?• routers may deliver datagrams destined to an anycast address

to any one of the group• work around: only first packet (SYN) uses anycast address as

dest. addr; subsequent packets use unicast address of the server


Why Isn’t Multicast Pervasive?

• (Reasonably) Sound technology– But fairly complex, with several versions of protocols

• Implemented in most routers

• Used by some enterprises

• But not deployed/used on public Internet


Possible Explanations• Lack of demand (up until now?)• Lack of membership/sender control

– Many services need it: who can join, who cannot– Also needed for billing, etc. – Hard to implement sender control:

• Can be subject to (and used to amplify) DoS attack

• Multicast address scarcity– global allocation required

• Inter-domain issues:– Violates current ISP settlement model– No incentive for ISPs to enable multicast

• ….


Design Alternative: Single-Source Multicast

• Multicast Channel Concept:– Each multicast group associated with a source– Defined by [source, multicast address]

• Implications:– Each group has only one source, address allocation non-issue!

• each source can allocate 16 millions “channels”• Sender control easy!

– Forwarding using source and dest. (multicast) IP addresses

• Use Reverse Path Multicast algorithm• Add a counting mechanism for estimating group size

– Use a recursive CountQuery message

• Use app-level relays to for multiple sources


Other Issues w/ Multicast

• FEC can help, but isn’t perfect

• Must have retransmissions

• But sender can’t keep state about each receiver– Has to be told when someone needs a packet

For example: How to make multicast reliable?(a la TCP to IP unicast)


SRM Design Approach• Let receivers detect lost packets

– By holes in sequence numbers

• They send NACK when loss is detected• Any node can respond to NACK• NACK/Response implosion averted through

suppression– Send NACKs at random times– If hear NACK for same data, reset NACK timer– If node has data, it resends it, using similar

randomized algorithm


• Chosen from the uniform distribution on

– A: node that lost the packet– S: source– C1,C2 : algorithm parameters– dS,A : latency between S and A– i : iteration of repair request tries seen

• Algorithm– Detect loss set timer– Receive request for same data cancel timer, set new timer– Timer expires send repair request

Repair Request Timer Randomization

])(,[2 ,21,1 ASASi dCCdC


Adaptive Timers• C and D parameters depends on topology and

congestion choose adaptively• After sending a request:

– Decrease start of request timer interval• Before each new request timer is set:

– If requests sent in previous rounds, and any dup requests were from further away:

• Decrease request timer interval– Else if average dup requests high:

• Increase request timer interval– Else if average dup requests low and average request

delay too high:• Decrease request timer interval


Local Recovery

• Some groups are very large with low loss correlation between nodes– Multicasting requests and repairs to entire group wastes

bandwidth

• Separate recovery multicast groups– e.g. hash sequence number to multicast group address– only nodes experiencing loss join group– recovery delay sensitive to join latency

• TTL-based scoping– send request/repair with a limited TTL– how to set TTL to get to a host that can retransmit– how to make sure retransmission reaches every host that heard

request


Discussion• Does multicast belong in the network layer?

– Why not implemented by end hosts?

• How important is economic analysis in protocol design?– Should the design drive economics, or the other way

around?• Multicast addresses are “flat”

– Doesn’t that make it hard for routers to scale?– Address allocation and aggregation?

• Should everything be multicast?

• What other delivery models are needed?


IP Anycast• Anycast: any member of a group• IP anycast service model

– Best effort, stateless– data sent to an anycast group is delivered any member

(preferably only one) of the group

• IP anycast address: represent a virtual host – use unicast IP addresses: assigned to multiple hosts

• advertise as host-specific route with a network– e.g., 128.101.35.34/32

• can be aggregated outside a network: 128.101.0.0/16• what about “global” anycast addresses, e.g., well-

known anycast address for autoconfiguration of DNS servers?

• hosts can’t distinguish unicast and anycast addresses!– reserve special IP address allocation for anycast

• either within a network address prefix – but won’t work for global IP anycast

• or a global IP address space– not aggregatable, scalability concerns


IP Anycast within a Network• Hosts that are members of an anycast group

– advertise membership to routers (e.g., via IGMP)– anycast address to MAC address translation:

• one option: map anycast IP address to unicast MAC address using ARP

– disadvantage? • another option: use link-level (e.g., Ethernet) multicast

MAC address, and map anycast IP address to it– hosts of anycast group on LAN configured as members of

link-level multicast group

• Sending hosts: anycast IP addr same as unicast addr• Role of routers

– listen to membership advertisement and keep track of anycast groups on subnets

– advertise anycast groups as host-specific routes • using unicast routing protocols (e.g., RIP or OSPF) • need some special care

– Forwarding: decide what next-hop to forward anycast datagram

• may need to map link- level multicast MAC address

internet design case study: ip multicast

Documents