CCTCP: A Scalable Receiver-driven Congestion Control Protocol for - PowerPoint PPT Presentation

CCTCP: A Scalable Receiver-driven Congestion Control Protocol for Content Centric Networking Lorenzo Saino, Cosmin Cocora and George Pavlou Communications and Information Systems Group Department of Electrical and Electronics Engineering University College London {l.saino,c.cocora,g.pavlou}@ee.ucl.ac.uk

Content Centric Networking (CCN) ◮ Content Centric Networking (CCN) 1 is a recently proposed Internet architecture which shifts the main network abstraction from node identifiers to location-agnostic content identifiers. ◮ In CCN, content objects are partitioned into addressable chunks which can be contained within a single packet and reactively cached at any router along a path. ◮ Two network primitives: ◮ Interest packets, which are routed according to the identifier of the requested content towards the closest available copy ◮ Data packets, which deliver the requested content chunk in response to an Interest packet 1 V. Jacobson, D. Smetters, J. Thornton, M. Plass, N. Briggs, and R. Braynard. Networking named content. ACM CoNEXT 2009

Transport issues in CCN Since contents are cached with a packet-level granularity, chunks may be served by different network nodes when retrieving an entire content object. This makes TCP-based implicit-feedback congestion control mechanisms inefficient: ◮ Out-of-order delivery or variations in inter-arrival times may be caused by adjacent chunks being served by different caches rather than congestion. ◮ RTO estimation is unreliable because of greater RTT variability caused by frequently changing chunk sources.

Proposed transport protocols for CCN Currently proposed transport protocols can be categorized in: ◮ Receiver-driven ◮ Control loop in the receiver, stateless routers ◮ Proposals: ICTP 2 , ICP 3 , ConTug 4 ◮ Hop-by-hop ◮ Control loop in the routers which need to keep per-flow state ◮ Possibility to control misbehaving receivers ◮ Proposals: HoBHIS 5 HR-ICP 6 2 S. Salsano, A. Detti, M. Cancellieri, M. Pomposini, and N. Blefari-Melazzi, ”Transport-layer issues in information centric networks”, ICN workshop, ACM SIGCOMM 2012 3 G. Carofiglio, M. Gallo, and L. Muscariello, ”ICP: Design and evaluation of an interest control protocol for content-centric networking”, NOMEN workshop, IEEE INFOCOM 2012 4 S. Arianfar, P. Nikander, L. Eggert, and J. Ott, ”Contug: A receiver-driven transport protocol for content-centric networks”, in IEEE ICNP 2010 (Poster session) 5 N. Rozhnova and S. Fdida, ”An effective hop-by-hop interest shaping mechanism for CCN communications, NOMEN workshop, IEEE INFOCOM 2012 6 G. Carofiglio and L. Muscariello, Joint hop-by-hop and receiver-driven interest control protocol for content-centric networks, ICN workshop, ACM SIGCOMM 2012

Content Centric TCP (CCTCP) Design objectives and assumptions Design Objectives ◮ Incremental deployability ◮ Scalability to Internet scale ◮ Fairness among CCTCP flows and with legacy TCP flows ◮ Independence from caching policies and caches location Assumptions ◮ Interest and Data packets follow the same path in opposite directions at the CCN layer ◮ Interest packets are routed to the original source. A cached copy is served only if on the path between receiver and original source

Content Centric TCP (CCTCP) Protocol overview ◮ Core features: ◮ Receiver-driven, window-based, implicit-feedback congestion control ◮ Slow start and congestion avoidance phases based on TCP New Reno ◮ Only timeout expiration used as signal of congestion. No window size drop on out-of-order arrivals. ◮ Fast recovery at timeout expiration ◮ CCTCP keeps one retransmission timeout and one congestion window per each expected source of chunks ◮ Expected source is predicted before sending an Interest packet, thanks to the anticipated interests mechanism

Anticipated interests mechanism B A RCV R k R n T I I N T E R E S T ( A anticipated interest: B ) T I (k) I N T E R E S T ( A B: Router = K, TS = T I (K) ) T D - T I T D (k) - T I (k) DATA(A) K ) T ( = S T I K , = r t e o u R T D (k) B : DATA(A) ) ( K T - K ) I ( = T ∆ T D K , = e r T D u t R o B : RTT ( k ) = ( T D − T I ) − ( T D ( k ) − T I ( k ))

Window and timeout update algorithms ◮ Timeout calculated according to RTO k-1 RTO k RTO k+1 RTO s CWND k-1 CWND k CWND k+1 CWND s TCP’s Jacobson algorithm RCV k-1 k k+1 s ◮ At each timeout expiration SLOW CONGESTION RTO ( i + 1) ← − 2 × RTO ( i ) and START AVOIDANCE CWND ( i + 1) ← − CWND ( i ) / 2 FLOW PROGRESSION ◮ At each Data packet reception, CWND ( k ) SRTT ( k ) ≥ CWND ( k + x ) rate is increased for all caches SRTT ( k + x ) between the receiver and the expected source RTO ( k ) ≤ RTO ( k + x ) ◮ At each timeout expiration, rate is ∀ x ∈ [1 , s − k ] increased for the expected source and all caches beyond it

Pacing of Interest packets ◮ Data packets are larger than Interest packets and therefore RCV SRC also their T tx is larger ◮ If Interest packets are sent in a INTEREST burts, Data packets sent in response to Interests at the end ) 1 ( T T ) 2 of the burst will experience R ( T T ) 3 DATA R ( longer RTT T T R ◮ Solution: pace the sending of Interest packets over an RTT period: T send ( i ) = SRTT ( i ) CWND ( i )

Performance evaluation Simulation scenario S R 10Mbps 10Mbps R S 40Mbps 40Mbps R C S C C C R S R S ◮ Metric: Flow Completion Time (FCT) ◮ Variable conditions: ◮ Cache sizes: 0 . 02 P and 0 . 08 P ◮ Content popularity α : 0.64, 1.03 ◮ Caching policies: ALL+LRU, ALL+LFU, RAND+LRU ◮ Comparison with ICP and CCTCP w/o Anticipated Interests

Performance evaluation Flow Competion Time (FCT), cumulative probability 1.0 1.0 0.8 0.8 0.6 0.6 CDF CDF 0.4 0.4 CCTCP, C=0.02P CCTCP, C=0.02P CCTCP, C=0.08P CCTCP, C=0.08P CCTCP w/o AI, C=0.02P CCTCP w/o AI, C=0.02P 0.2 0.2 CCTCP w/o AI, C=0.08P CCTCP w/o AI, C=0.08P ICP, C=0.02P ICP, C=0.02P ICP, C=0.08P ICP, C=0.08P 0.0 0.0 0 5 10 15 20 25 30 0 5 10 15 20 25 30 35 Flow Completion Time (sec) Flow Completion Time (sec) (a) ALL+LRU, α = 0 . 64 (b) ALL+LRU, α = 1 . 03 1.0 1.0 0.8 0.8 0.6 0.6 CDF CDF 0.4 0.4 CCTCP, C=0.02P CCTCP, C=0.02P CCTCP, C=0.08P CCTCP, C=0.08P CCTCP w/o AI, C=0.02P CCTCP w/o AI, C=0.02P 0.2 0.2 CCTCP w/o AI, C=0.08P CCTCP w/o AI, C=0.08P ICP, C=0.02P ICP, C=0.02P ICP, C=0.08P ICP, C=0.08P 0.0 0.0 0 5 10 15 20 25 30 0 5 10 15 20 25 30 35 Flow Completion Time (sec) Flow Completion Time (sec) (c) RAND+LRU α = 0 . 64 (d) RAND+LRU, α = 1 . 03

Performance evaluation Impact of inter-cache distance 1.0 0.8 0.6 CDF 0.4 CCTCP, 1 ms CCTCP, 5 ms CCTCP w/o AI, 1 ms 0.2 CCTCP w/o AI, 5 ms ICP, 1 ms ICP, 5 ms 0.0 0 5 10 15 20 25 30 Flow Completion Time (sec)

Conclusions and future work ◮ Previously proposed receiver-driven transport protocols for CCN cannot reliably identify congestion by keeping a single timeout value. ◮ Hop-by-hop protocols can cope with with source variability but at the cost of requiring each router to keep per-flow state. ◮ CCTCP considerably outperforms other receiver-driven proposals and the performance gain provided by anticipated interests is substantial. ◮ Results presented here have been gathered using an unoptimized algorithm for selecting anticipated interests chunks set. Optimized algorithms can further improve performance.