A Comparison of Application-Level and Router-Assisted Hierarchical Schemes for Reliable Multicast Pavlin Radoslavov Christos Papadopoulos Ramesh Govindan.

A Comparison of Application-Level and Router-Assisted Hierarchical Schemes for

Reliable Multicast

Pavlin RadoslavovChristos Papadopoulos

Ramesh GovindanDeborah Estrin

Reviewer: Jing Lu, Qian Wan

CS770x

Outline

• Introduction– ALH: RMTP

– RAH: LMS

• Metric Space

• Analysis Using k-ARY Trees

• Simulation Results

• Conclusion

IP Multicast

• Send packet from a source to the members of a multicast group.– Class D IP addresses (250 million)

– IGMP & MOSPF

– Best-effort packet forwarding

• Applications: multimedia, teleconferencing, distributed computing, etc.

Reliable IP Multicast

• Scalability issues:– Implosion: redundant messages triggered by packet loss

– Exposure: redundant retransmissions to receivers who haven't experienced loss

• Long recovery latency

• Hierarchical data recovery schemes:– ALH (Application-Level Hierarchical): End systems assist in

hierarchy creation and maintainance.• RMTP

– RAH (Router-Assisted Hierarchical): Routers assistance• LMS

RMTP Data Recovery

• Static hierarchical scheme– Designated Receivers (DRs) are chosen statically

– A receiver dynamically chooses a closest DR as its Ack and retransmission processor

– A DR collects Nack from its local group members and retransmits packet within the group using unicast/multicast

– A DR emits its own Nack to its parent DR in the upper hierarchy

– Sender deals with Nacks from DRs at the top level hierarchy

ALH Data Recovery

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

• Optimal Hierarchy

ALH Data Recovery

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

• Optimal Hierarchy

ALH Data Recovery

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

• Sub-optimal Hierarchy

ALH Data Recovery

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

• Sub-optimal Hierarchy

Heuristic Dynamic Hierarchy Creation in ALH

• Each receiver obtains distance info to each other

• Dynamically create the hierarchy from bottom-up:– Initially all receivers are eligible to become parents

– A fraction (fracpc) of receivers with the smallest sum of distances becomes parents.

– Receivers that are not elected choose the closest parent as its parent.

– Repeat the selection process among receivers chosen from the previous iteration until the number of receivers left <= 1/fracpc, so their parent is the sender itself.

LMS Data Recovery

• LMS extends router forwarding

• Enhance routers to:– Replier selection

– Forward Nacks to replier and discover root of loss subtree

– Perform DMCAST

LMS Replier Selection

• Router state per-source tree:– Upstream link

– List of downstream links

– Replier link id

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

LMS Nack Forwarding

LMS router handles Nacks [1]

LMS DMCAST

• DMCAST:– Replier encapsulates a multicast packet into a unicast packet and

sends to the turning-point router

– LMS router decapsulates and multicasts it on the specified link interfaces

LMS Enhanced Two-Step DMCAST

• Nack from a downstream replier specifies reply should be unicast back to it rather than to its turning point

• Replier then performs DMCAST when necessary

Summary of ALH and RAH

ALH RAH

Automatic creation of data recovery hierarchy

End-to-end mechanism and heuristic algorithm

Router selects the closest downstream receiver as replier

Retransmission Parent unicasts/multicasts recovery data to its group members

Replier unicasts recovery data to turning-point router, router multicasts it directly on specified links

• RAH is finer-grained with many more “internal nodes”• RAH is more congruent to the underlying multicast tree• RAH doesn’t have explicit group concept, so it is easily adaptive to membership change; membership maintenance cost is minimal

Metric Space

• Data Recovery Latency

• Receiver Exposure

• Data Traffic Overhead

• Control Traffic Overhead

Data Recovery Latency

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

Loss Rcvs lat RTT

Rx2 6 8

Rx3 8 10

Rx4 8 10

Rx5 8 10

Rx6 8 10

NormLat 0.79

Receiver Exposure

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

Loss Rcvs Exposure

Rx2 0

Rx3 0

Rx4 0

Rx5 0

Rx6 0

NormExp 0

Data Traffic Overhead

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

Loss Rcvs Data Subtree

Rx2 3 8

Rx3, Rx4, Rx5, Rx6

7

NormDataOverhead

1.25

Control Traffic Overhead

sender

R1

R3

R2

R4

Rx1Rx7

Rx8

Rx3 Rx4 Rx5 Rx6

Rx2

Loss Rcvs Control Subtree

Rx2 3 8

Rx3 3

Rx4 3

Rx5 3

Rx6 3

NormLat 1.875

Analysis using k-ARY Tree

• Purpose:− Gain initial understanding of the scalability of the ALH and RAH schemes

• Parameters:− k, L− q: fraction of leaf nodes that are receivers is 1/kq-1

• Assumptions:− Each parent (ALH) has k-1 children.− Single link loss and average per link-loss across all links

Analysis using k-ARY Tree

• ALH

• RAH

Control Overhead Analysis

L = 10

Data Overhead Analysis

L = 10

• RAH is slightly better than ALH• In some cases, RAH replier multicast data to all receivers within a subtree• ALH has to perform multiple multicasts within local groups

Data Recovery Latency Analysis

Data Recovery Latency Analysis

L = 10