190 likes | 204 Views
Congestion Control for Multipoint Communications in ATM Networks. U.T. Nguyen Dept. of Computer Science and Engineering York University. Presentation Contents. Motivations Outline of research topics and contributions Explicit rate control for ABR multipoint-to-point
E N D
Congestion Controlfor Multipoint Communicationsin ATM Networks U.T. Nguyen Dept. of Computer Science and Engineering York University Department of Computer Science
Presentation Contents • Motivations • Outline of research topics and contributions • Explicit rate control for ABR multipoint-to-point • Explicit rate control for ABR multipoint-to-multipoint • Reliable multicast at ATM layer motivations, proposed solution, contributions • Conclusion and future work
Congestion Control Is Essential! • For any type of network • Traffic demands are unpredictable • Congestion occurs when rate of arrival exceeds rate of processing/transmitting • Consequences of congestion: – collapse of network throughput – long delay experienced by users – data loss
Even More Essential ... • In ATM networks • several service classes with different QoS guarantees • high bit rates • When congestion occurs in ATM networks, • data queue up quickly at switches • users suffer long delay, excessive cell loss • network can’t deliver QoS guarantees • For multipoint communication (and more challenging!) • coordination of multiple senders and receivers • much higher bandwidth consumption compared with unicast connections
Our Research Focus Congestion Control for Multipoint Communications in ATM Networks
ATM Congestion Control There are 3 mechanisms: • Cell discarding • When buffer occupancy reaches a threshold • Applies to all service classes: UBR ABR VBR/CBR • Explicit rate control for ABR services • Available bandwidth after VBR/CBR usage is divided evenly among ABR connections • RM cells convey the allowed cell rate info to ABR sources • Explicit forward congestion indication • Switches set EFCI bit when experiencing congestion • Actions to control congestion are application-dependent
Contributions of the Research • Explicit rate (ER) control for ABR services We propose effective and flexible algorithms for • Explicit rate control for ABR MP-P • Explicit rate control for ABR MP-MP • Cell discarding • Existing reliable multicast protocols are at transport layer • One cell dropped the whole packet retransmitted • Multicast: many copies of a packet are re-sent redundantly • Consequences: increase of congestion level; high recovery latency • We propose: loss detection/retransmission at ATM layer, cell level Reliable multicast at the ATM layer that reduces congestion and minimizes recovery latency
Motivations Several bandwidth allocation definitions (BADs) Source-based VC/source-based Weighted-source-based Flow-based VC/flow-based Weighted-flow-based No consensus on which BAD is best Existing explicit rate control schemes for ABR MP-P support 1 or 2 BADs, or are not effective: Ren, Siu and Suzuki: source-based incorrect w/ VC merge Moh and Chen source-based weighted-source-based incorrect in some cases Fahmy, et al.: source-based only Explicit Rate Control for ABR Multipoint-to-point Services
Feasibility of BADs • For simplicity: • source-based feasible • VC/source-based: MP sources are penalized for sharing resources • For fairness of pricing: • weighted-source-based is the best most feasible • Extending weighted-source-based definition even further: • every source, unicast or MP, has its own weight • more overheads; more complex to implement • ABR is a best-effort service not worth the increased cost • BADs based on flows: • sources merging many times may get a smaller share • advantage: simpler to implement than BADs based on sources
Advantages of NK Algorithm • Flexible • supporting both VP merge and VC merge • supporting a large set of BADs • Minimal overhead • overhead for implementing weighted-source-based BAD • accommodating other BADs with no overhead increase • easy setting using only a few software parameters • Scalable and efficient • one FRM is replied by exactly one BRM • running time of NK algorithm: O(B) • Effective allocation, fast convergence, robustness • by design (e.g., using fair share, computing ER in both directions) • confirmed by simulation results
Experimental Results • Simulation • ATM network simulator version 4.1 by NIST • code added to implement MP-P routing, NK and FJ algorithms • Measures • ACRs of sources: effectiveness of allocation, convergence time • queue lengths at switches: cell loss, stability • Performance of NK algorithm • under normal condition, heavy load, with join/leave operations • correct allocation, fast convergence, stable after convergence • minimal rate oscillation during transient periods • adjusting well to heavy traffic • Comparing NK and FJ algorithms: NK algorithm performs better • faster convergence • shorter queue lengths during transient periods less cell loss • much smaller rate fluctuations
Explicit Rate Control for ABR Multipoint-to-multipoint Services Motivations • One single VC for the entire MP-MP group • Several bandwidth allocation definitions (BADs) • No consensus on which BAD is best • Existing explicit rate control schemes for ABR MP-MP support only source-based allocation, or are not effective or scalable: • Ren, Siu and Suzuki • source-based allocation • incorrect calculation with VC merge • Pao • source-based allocation only • Cavendish and Gerla • MP-MP group with n members = n P-MP connections not scalable
Advantages of MPER Algorithm • Flexible • supporting both VP merge and VC merge • implementing a large set of BADs • Minimal overhead • overhead for implementing weighted-source-based BAD • easy setting using only a few software parameters • Scalable and efficient • a MP-MP group is supported by a single VC • one FRM is replied by exactly one BRM • running time of MPER algorithm: O(B) • Feedback consolidation at branch points • minimizing rate fluctuation and cell loss rate • maintaining stability after convergence • Effective allocation, fast convergence, robustness • by design (e.g., using fair share, computing ER in both directions) • confirmed by simulation results
Experimental Results • Simulation • ATM network simulator version 4.1 by NIST • Code added to implement MP-MP routing and MPER algorithm • Measures • ACRs of sources: effectiveness of allocation, convergence time • Queue lengths at switches: cell loss, stability • Performance of MPER algorithm • under normal condition and heavy load • correct allocation, fast convergence • minimal rate oscillation during transient periods • stable after convergence thanks to feedback consolidation • adjusting well to high traffic demands
Reliable Multicast at the ATM Layer Motivations • Existing reliable multicast protocols are at transport layer • one cell dropped the whole packet retransmitted • multicast: many copies of a packet are retransmitted redundantly • Consequences: • increase of congestion level • high recovery latency Our Objective • A reliable multicast protocol that • minimizes redundant retransmission (exposure) • offers low recovery latency, scalability • We propose: • loss detection/retransmission done at ATM layer • unit of recovery: ATM cells, not packets
REMAT Protocol • Adding 2-byte sequence numbers to ATM cells • sufficient to support applications requiring both reliability and real-time delivery • incurring less overhead than AAL3/4 (2 bytes versus 4 bytes) • Request/retransmission algorithm • uses the same ATM routing tree • local recovery performed by retransmission servers: • child-parent request/retransmission • scalable; low recovery latency • no implosion of requests or repair data • exposure, if any, is limited to one level of children • Message stability detection: a feedback control mechanism • for managing buffers • that can be used for rate-based congestion control
Experimental Results • Simulation • ATM network simulator version 4.1 by NIST • code added to implement P-MP routing, REMAT and LSM protocols • Comparing • ATM-layer recovery: REMAT protocol • transport-layer recovery: LSM-like protocol • Measures • average retransmission delay • recovery latency • the faster packets are cleared, the better to alleviate congestion • connection throughput P= S/T • performance gain/loss • Results: ATM-layer recovery offers • much lower average retransmission delay in all cases • better connection throughput under moderate to heavy congestion
Summary • Congestion control is crucial to • high speed networks like ATM • multipoint communications • delivering QoS guarantees • We propose flexible and effective algorithms for • Explicite rate control for ABR multipoint-to-point services • Explicite rate control for ABR multipoint-to-multipoint services • fast convergence, minimal rate fluctuation • adapting well to heavy traffic demands • minimal implementation overhead • We propose REMAT protocol that performs loss recovery at the ATM layer • scalable: no implosion, minimal exposure • reducing unnecessary retransmission of transport-layer reliable multicast • much lower recovery latency than that of transport-layer reliable multicast
Future Work • Defining MCRs and PCRs of multipoint connections (per VC or per source?). How should the rate calculation algorithm and connection admission protocol be modified? • Supporting heterogeneous receivers in an MP-MP group using layered data approach • Developing the message stability mechanism of REMAT protocol into a rate control scheme for flow/congestion control