Presentation slides prepared by Ramakrishnan.V

LMS: A Router Assisted Scheme for Reliable Multicast

Christos Papadopoulos, University of Southern California

Guru Parulkar, Washington University

George Varghese, University of California at San Diego

Introduction

Internet architecture is largely responsible for the undisputed success of the Internet today.

Unfortunately, the end-to-end model which was so successful in the unicast case has proven much harder to apply to multicast.

LMS facilitates very efficient solutions (compared to pure end-to-end schemes) to problems like scalable reliable multicast.

Difficulties - Reliable Multicast

Implosion

Host Exposure

Recovery latency

Adaptability to dynamicity

Key Feature of LMS (Light-Weight Multicast Service)

Separation of forwarding and error control Forwarding (by routers) Error control (by receivers – end to end)

No packet storing or processing at routers.

Core Ideas

Each router selects a replier (surrogate). Routers steer requests to repliers. Routers help repliers multicast replies to loss

sub-tree. LMS restricts the scope of forwarding.

LMS achieves the efficiency of the heavy-weight model, but without the weight.

LMS Model

R

Control messages

Router storespackets, receivesNACKs and sendsretransmissions

Heavy-weight model

R

Router chooses a receiveras a surrogate.

Router relays messagesfrom surrogate to theSub-tree.

Receiver actingas a surrogate

LMS

Router steers all controlmessages to surrogate.

Idealized Recovery Scenario

LMS: Concepts

Replier Provides Retransmission Receiver volunteered to

answer requests Turning point

Where requests start to move downstream

Directed Multicast Multicast to a sub-tree

SS

RR

RR

AA AA AA AA

AA

PP

PP

PP Replier

Replier link

Replier link

Selecting a Replier (surrogate)

If the router has two or more downstream links it selects one as the replier link.

If the router has only one downstream link that becomes the replier link by default.

If the source is directly attached to the router the source becomes the replier.

Replier Selection (contd…)

Receivers express desire by piggy backing information on the join request.

Receivers communicate a cost of their appropriateness as repliers.

Selection based on the advertised cost.

Steering Messages to Repliers

Each request is multicast, which keeps receiver actions simple.

Hop-by-Hop forwarding requires routers to examine each request, which is done via the IP Router Alert option, included in every request.

Request Handling at the Routers

LMS avoids request implosion because each router allows only one request to escape upstream - the one coming from the replier link.

Directed Multicast (DMCAST)

Its purpose is to enable fine-grain multicast to eliminate exposure.

LMS: Request forwarding

Multicast to the group.

If a request reaches a turning point, it’s forwarded towards the replier.

No request suppression or merging, but scope of requests is limited.

SS

R2R2

R1R1

AA AA AA AA

AA

P2P2

P1P1

LMS: Reply forwarding At turning point,

<router_addr:link_id> is added into request packet.

Replier includes it into its retransmission packet.

Routers need not remember anything about requests as they pass through.

Routers are not even aware that these recovery messages.

SS

R2R2

R1R1

AA AA AA AA

AA

P2P2

P1P1

turning point

LMS Concepts review

Concepts Replier selection. Steering of requests to repliers. Establishing turning points. Directed multicast.

These concepts work together to enable receivers to construct an efficient recovery mechanism.

Problem: Exposure

Loss at the replier link may result in duplicates Mitigation: Using the cost field to select a replier that

advertises the least loss.

Simulations

Using 3 Topologies Binary Trees Random Topology Transit-stub topology

Using 100’s of nodes and tens of topologies.

Compare with PGM and SRM.

Simulation Parameters

Loss at the source Test control of NACK implosion.

Loss at the receiver Test control of Exposure.

Loss at each Link Working in random loss scenario.

Simulation Summary

LMS performance improves as the group gets larger because more helpers.

LMS and PGM to perform much better than SRM.

Comparing LMS and PGM LMS is much simpler to implement, and

performance is on par with PGM. Significantly lower recovery latency, while

trading very little in terms of exposure.

Late Requests lead to ambiguity

Replier Failure

Selecting Repliers in a LAN

So far we have assumed that there was only one receiver at each link.

Receivers on a LAN use a simple election mechanism to elect a replier and therefore make a LAN appear as having a single receiver.

Routers with a Large Fan-out

Proxy Directed Multicast

The request may arrive after the buffers at the replier have been purged.

Note that once a request passes the turning point it contains enough information to uniquely identify the sub-tree that requires the retransmission.

Replier can forward the request to another member (LMS again).

Pathological Topologies

Incremental Deployment

Use Source Path Messages (SPMs)to create an overlay.

Incremental deployment has an effect on exposure.

Other Applications - ANYCAST

Grouping servers in a well-known multicast group. Servers tell the routers to advertise the existence of a

replier (server) in all links which ensures that routers find the nearest server in any direction.

Summary

LMS is simple to implement, does not violate the end-to-end principle.

Forwarding and error control are two clearly separable components, and great benefits can be realized by decoupling and placing each one where it is more beneficial.