Transport Control Protocol over Optical Burst Switched Networks

Transcription

1 Transport Control Protocol over Optical Burst Switched Networks Basem Shihada Research Associate Electrical & Computer Engineering University of Waterloo Seminar, July 3 rd 2008

2 Seminar Outline Introduction to OBS & TCP What is the problem with running TCP over OBS? Why the TCP over OBS problem is important? How this problem is solved? What are the contributions? Ongoing & Future work Conclusion 2

3 Application Demands Applications Requirement Optical Transport Paradigm Voice Over IP Streaming Video Grid Computing Multimedia Data High-speed Data Transmission Low Loss Quality of Service Optical Circuit Switching Optical Packet Switching Optical Burst Switching OBS is promising due to: no optical buffer required and dynamic bandwidth allocation 3

4 OBS Network Architecture Output Traffic Edge Node Edge Node Input Traffic IP Edge Node Burst Assembly Routing Header Offset Time Burst Core Node Signaling Scheduling 4

5 Performance Measurement in OBS Burst loss Due to burst contentions at intermediate nodes Burst Header Burst a b a b Wavelength Time Burst delay t 1 t 2 t 3 Burst assembly delay Offset time One-way propagation delay Burst Contention 5

6 Key Points! Optical Burst Switching (OBS) Networks Shift control complexity from optical to electrical layer Lower the switching granularity Dynamic bandwidth efficient All-optical bufferless Highly synchronized Suffers from random burst contention losses 6

7 Transport Control Protocol (TCP) Contributes to the success of the Internet through: Reliable data transfer Self regulated Congestion tolerant 80-90% of the Internet traffic comes from TCP sources Implements four major functional modules: slow start, congestion avoidance, fast retransmission and fast recovery Additively increases the sending rate to probe the available bandwidth and Multiplicatively decreases the sending rate at the occurrence of segment loss by: half in dropping-based TCP quarter in delay-based TCP β in TCP which uses the Generalized AIMD (GAIMD) 7

8 Problems of Dropping-based TCP over OBS Dropping-based TCP (e.g., Reno) can NOT distinguish whether packet loss is due to congestion or a random burst contention What happens for Reno over barebone OBS and OBS with burst retransmission networks? Case 1: low traffic load Fast Reno flows fall into false congestion detection Medium to slow Reno flows trigger unnecessary TDs The loss of Acks damages Reno s self clocking mechanism Case 2: high traffic load Higher congestion and more burst losses Reno suffers from more frequent losses which significantly reduces the throughput 8

9 TCP & Random Burst Loss Multiple TCP sources unnecessarily reduce their transmission rate due to single burst loss Single TCP source is fatally affected by a single burst loss (concentrated loss) Waste of routing, assembly, and signalling efforts performed at both IP-access and OBS networks Multiple simultaneous burst losses can cause a network-wide loss in throughput (global synchronization problem) Significantly affect the long-lasting high-bandwidth flows Need an effective mechanism to solve the false congestiondetection problem 1 1. B. Shihada, P-H. Ho, F. Hou, X. Jiang, S. Horiguchi, M. Guo, & H. Mouftah, "BAIMD: A Responsive Rate Control for TCP over Optical Burst Switched (OBS) Networks", IEEE International Conference on Communication (ICC), Istanbul, Turkey, June,

10 Problems of Delay-based TCP over OBS Delay-based TCP (e.g., Vegas) can NOT detect network congestion in a barebone OBS network, since RTT varies little!!! What happens for Vegas over an OBS network with burst retransmission or burst deflection? Case 1: low traffic load Vegas detects sudden RTT increase due to retransmission or deflection Results in unnecessary reduction of the cwnd size! Case 2: high traffic load Higher congestion and longer delay Vegas detects more frequent RTT increase, which significantly reduces the throughput 10

11 TCP & Burst Delay Assembly delay penalty Typical assembly time ranges from few tens of μs to hundreds of ms (i.e., sometimes the burst assembly delay is few times larger than the link propagation delay!!) Takes place at both network edges and affect the data segments along with the Ack packets Reduces the link utilization since it increases the RTT and the RTO Introduces buffering constrains at the edge nodes Retransmission penalty TCP is busy retransmitting lost packets in a burst that contain many packets Few new packets can be sent due to prolong retransmission period 11

12 TCP & OBS with BCR OBS network with burst contention resolution (BCR) schemes Burst deflection Burst retransmission Burst optical buffering using Fiber Delay Lines (FDL) Burst segmentation Reduces burst loss probability Introduces additional delay Potential cause of burst out-of-order delivery TCP suffers from Significant throughput reduction due to enlarging RTT Unnecessary data retransmissions due to false TDs Unnecessary TCP transmission rate reduction due to False RTOs 12

13 Statistical AIMD Goal: to design a proper TCP congestion avoidance mechanism that is suitable for detecting burst losses due to random contention in OBS networks 1,2 Propose a framework for determining network congestion based on statistical RTT information and confidence Propose a modified congestion control algorithm by manipulating a factor 0.5<β<1 to multiplicatively decrease the congestion window Provide an analytical model for SAIMD over OBS Simulate SAIMD over barebone OBS, OBS with burst retransmission, OBS with burst deflection 1. B. Shihada, P-H. Ho, & Q. Zhang, SAIMD: A Congestion Detection Scheme for TCP over OBS Networks, accepted in IEEE/OSA Journal of Lightwave Technology, B. Shihada, P-H. Ho, & Q. Zhang, A Novel Congestion Detection Scheme for TCP over OBS Networks, 50th IEEE Global Communications Conference (GLOBECOM), Washington DC, USA, November,

14 TCP with Burst Contention-Loss Notification Goal: to design a proper TCP congestion control mechanism called (TCP-BCL) 1 that distinguishes different burst losses relying on the explicit notification from the OBS network 2 Propose a framework for identifying burst congestion based on explicit information from the OBS edge node Modifies TCP congestion control algorithm by manipulating a factor 0.5<β<1 to multiplicatively decrease the congestion window as per the received contention loss signal Provide an analytical model for TCP-BCL over OBS Simulate TCP-BCL over barebone OBS and OBS with burst retransmission 1. B. Shihada, and P-H. Ho, "A Novel TCP with Dynamic Burst-Contention Loss Notification over OBS Networks", Elservier Journal of Computer Networks,Vol. 52/2, pp , B. Shihada, P-H. Ho, & Q. Zhang, "TCP-ENG: Dynamic Explicit Congestion Notification for TCP over OBS Networks", 16th IEEE International Conference on Computer Communications and Networks (ICCCN 07), Hawaii, USA, August,

15 Threshold-Based TCP Vegas Goal: to understand the behavior of delay-based TCP over OBS networks, so as to design a proper TCP congestion avoidance mechanism for OBS networks Design and simulate a Threshold-based TCP Vegas over barebone OBS, OBS with burst retransmission and burst deflection 1 Model TCP Vegas behavior over barebone OBS and OBS with burst retransmission 2 Model Threshold-based TCP Vegas behavior over barebone OBS and OBS with burst retransmission 3 1. B. Shihada, Q. Zhang, & P-H. Ho, "Threshold-based TCP Vegas over Optical Burst Switched Networks", 15th IEEE International Conference on Computer Communications and Networks (ICCCN'06), Virginia, USA, October B. Shihada, Q. Zhang, & P-H. Ho, "Performance Evaluation of TCP Vegas over Optical Burst Switched Networks", IEEE Broadnets, the 6th International Workshop on Optical Burst/Packet Switching (WOBS), San Jose, USA, October, B. Shihada, Q. Zhang, P-H. Ho, & J. Jue, A Novel Implementation of TCP Vegas for Optical Burst Switched Networks, Second Revision, IEEE Journal on Selected Areas in Communications (JSAC), April,

16 Contributions 1 Solution Category OBS Type Devices Involved Problems Addressed Schemes Notification Without Notification Barebone BCR TCP Sender OBS edge OBS core Random Burst Loss False TO Packet Reordering Burst AIMD TCP-BCL TCP-ENG Statistical AIMD Thresholdbased Vegas 1. B. Shihada & P-H. Ho, "Transmission Control Protocol (TCP) in Optical Burst Switched Networks: Issues, Approaches, and Challenges, IEEE Communications Surveys and Tutorials, Second Quarter, Vol. 10, No. 2,

17 Ongoing & Future Work Ongoing: Examine the effects of burst delay and losses on TCPs designed for Long-haul delay products: Fast TCP High-Speed TCP Binary Increase Control (BIC) Investigate the number of burst retransmission and deflection attempts on TCP and OBS load Future work: Introduce a cross-layer design for integrating TCP with the advantages of OBS with BCR Capture the characteristics of TCP behavior over heterogeneous network. 17

18 Conclusion Highlighted the problem of running TCP (dropping-based, delaybased) over barebone OBS and OBS with BCR networks Highlighted the significance of the problem in terms of the two major OBS performance measurements Random burst contention losses Various burst delays Introduced TCP congestion-control approaches to cope with the burst transmission behaviour Without explicit notifications Burst AIMD (BAIMD) Statistical AIMD (SAIMD) Threshold-based TCP Vegas With explicit notifications TCP with Burst Contention Loss Notifications (TCP-BCL) TCP with Explicit Notification GAIMD (TCP-ENG) 18

19 Questions & Discussion Transport Control Protocol over Optical Burst Switched Networks Basem Shihada Research Associate Electrical & Computer Engineering University of Waterloo 19

20 Backup Slides!

21 TCP over OBS Networks: Challenges and Approaches Basem Shihada University of Waterloo

22 TCP Congestion Control TCP congestion control approaches Dropping based, e.g. Reno, New Reno, SACK Delay based, e.g. Vegas Mix of dropping and delay, e.g. FAST TCP Explicit notification based, e.g. XCP, ELN, ECN Rate based, e.g. TCP Westwood, TCP Real Congestion control: AIMD window-based Additively increases the sending rate (1 segment per round) to probe the available bandwidth. Multiplicatively cuts the cwnd by half at the occurrence of segment loss in dropping-based TCPs. quarter at the of segment loss in delay-based TCPs. β in TCP which follows the Generalized AIMD (GAIMD). Implements four major functional modules: slow start, congestion avoidance, fast transmission and fast recovery. Maintain fairness among the co-existing flows. 22

23 Impact of Burst Delay on TCP Assembly delay penalty Typical assembly time is between few hundreds of ns to few hundreds of ms (i.e., sometimes the burst assembly delay is few times larger than the link propagation delay!!) Assembly delay takes place at both network edges and affects the data segments and the Ack packets. Reduces the link utilization since it increases the RTT and the RTO Introduces buffering constrains at the edge nodes Retransmission penalty TCP is busy retransmitting lost packets in a burst that contain many packets Few new packets can be sent due to prolong retransmission period Delay the first loss gain (DFL) Enlarges the transmission unit from packet to burst! The larger the burst size (# of assembled packets from the same source) the larger the throughput and the larger the DFL 23

24 OBS Taxonomy Barebone OBS Burst contention results into an immediate burst loss! OBS with burst contention resolution (BCR) Burst retransmission Burst deflection Burst segmentation Burst Buffering (FDL) Reduce burst loss probability Introduce additional delay 24

25 Methodologies Link-Layer solution Mechanisms undergoing in the OBS domain May not be sufficient; not adaptive to the TCP dynamics Congestion detection with explicit notification Can effectively solve the false congestion problem Cause signalling and nodal processing overhead Congestion detection without explicit notification TCP senders estimate/evaluate the OBS congestion status Less computation overhead 25

26 Taxonomy of TCP over OBS Link Layer Congestion Detection without explicit notifications Congestion Detection with explicit notifications Burstification Burst contention recovery Burst dropping policy Statistical AIMD Threshold-based TCP Vegas BTCP (BLE) BTCP BAIMD (BACK,BNAK) TCP-BCL 26

27 Link-Layer Solutions Burstification Process Adaptive burst assembly (AAP) Burst Contention Recovery Retransmission Deflection Optical Buffering Burst Segmentation TCP with burst Acknowledgment TCP Decoupling Retransmission-Count based Dropping Policy 27

28 TCP Congestion Detection with Explicit Notifications Burst TCP with burst Ack/Nack Partitions TCP rounds at the edge or core nodes Acks are received at the TCP sender before the actual segment delivery Nacks are triggered from the network core Complicates the functionality of the network nodes and introduces extra computation overhead. Violates the TCP semantics! TCP with burst contention loss (TCP-BCL) Combines GAIMD (α,β) with burst-contention notification from the edge nodes Uses burst-loss statistics to estimate the reason of the burst loss Reduces the core-network computation overhead! 28

29 TCP Congestion Detection without Explicit Notifications Burst TCP with Burst Length Estimation (BLE) Solves the problem of TCP false TOs Maintains a burst cwnd (burst_wd) Benefits from the TCP TO behavior and the number of segments sent The accuracy of estimation is a challenge! Burst AIMD (BAIMD) Estimates the network load to control the cwnd size Statistical AIMD (SAIMD) Uses RTT statistics Defines network congestion based on confidence Fails to work on networks with no RTT variation! Threshold-based TCP Vegas Reduces Vegas congestion reaction to non-congestion RTT increases Optimizing threshold is a challenge! 29

30 Open Research Problems Burst delivery fairness Merit-based channel allocation Batch scheduling algorithms Random early discard (RED) Slow convergence To increase the TCP cwnd from half to full utilization of 10Gbps with 1.5 Kbyte packets, we need 1 hour with 100ms RTT and p<10-9 Effect of burst delay and dropping on Fast TCP High-Speed TCP Binary Increase Control (BIC) XCP TCP performance evaluation Packet-oriented Fluid Models Very large number of TCP flows Poisson arrival of loss events Strong correlation between losses in one RTT Synchronization models 30

31 SAIMD: False Congestion-Detection Scheme for TCP over OBS Networks Basem Shihada University of Waterloo

32 TCP over OBS Burst contention occurs even at low traffic loads Bufferless One-way resource reservation signaling False congestion detection by the dropping-based TCP Contention resolution schemes Reduce burst loss probability Introduce additional delay False congestion detection by the delay-based TCP Burst loss and delay affects Single TCP (fast flows) Multiple TCPs (medium to slow flows) 32

33 False Congestion Detection in TCP over OBS When a burst containing segments from different TCP sources is dropped, multiple TCP sources will throttle back their transmission When a burst containing multiple segments from a single TCP flow is dropped, it can fatally affect that TCP flow transmission Multiple simultaneous burst drops in a congested network can cause a network-wide loss in throughput (global synchronization problem) May address a profound malicious impact on the long-lasting highbandwidth TCP flows Need to develop an effective approach to solve the false congestion-detection problem 33

34 Previous Methodologies Link-Layer solution Mechanisms undergoing in the OBS domain May not be sufficient; not adaptive to the TCP dynamics Congestion detection with explicit notification Can effectively solve the false congestion problem Cause signalling and nodal processing overhead Congestion detection without explicit notification TCP senders estimate/evaluate the OBS congestion status Less computation overhead 34

35 Statistical AIMD (SAIMD) Adapts the congestion detection without explicit notification Uses RTT to sense the network congestion Adopts the framework of GAIMD => ( ) instead of (1,0.5) Derive a histogram curve by the statistics of the long-term measured RTT The long-term RTT statistics represent the overall effect on the TCP sender due to network topology, routing strategy, and long-term traffic distribution In the modeling, we assume the spectrum to follow a Normal distribution 35

36 Statistical AIMD (SAIMD) (cont.) For every TCP segment loss 1. Derive avg_rtt_n as the short-term average RTT, and position it in the spectrum of RTT obtained from the long-term statistics 2. Obtain the confidence that the current network is in congestion state 3. Use confidence to determine a beta value for cwnd adjustment corresponding to the segment drop If the short-term RTT is similar to or even less than the long-term RTT, the segment loss event is more likely caused by random contention => cwnd is slightly cut 36

37 SAIMD Congestion Control using Confidence z 1, low confidence of congestion, z n, high confidence of congestion, u u ( ) 1 2( u u ) i 1 z 1 z i z n, potential for congestion, f ui n 1 37

38 Autocorrelation Goal: to determine a proper threshold for starting the short-term RTT statistics (N). Autocorrelation function: 1 N R(0, N ) RTT ( i) RTT ( i N ) N i 0 R(0,N) has a maximum value with N=0. SAIMD selects the N range such that R(0,N) = R(0,0). γ% 38

39 SAIMD Flowchart 39

40 Extreme Cases Case 1: SAIMD starts in a congested network The short and long term RTT statistics are very close. Hence The cwnd will not be reduced properly SAIMD eventually TO as a response to a packet loss and enters the slow start Case 2: SAIMD operates with no RTT variation Rare event, but possible in barebone OBS. The short and long term RTT statistics are also very close. Hence SAIMD uses TO as a reaction to network congestion 40

41 SAIMD Performance Modeling Goal: to understand the behavior of SAIMD over OBS networks Obtained: SAIMD performance model with low and high burst contention probability SAIMD behavior while collecting TDs. SAIMD behavior while triggering TOs. 41

42 Simulation and Analytical Results NSF network topology 8 wavelengths, each operates at 10Gbps (OC192) Burst timeout is 5msec TCP packet size is 1KB Burst offset time is 6µsec. LAUC-VF is implemented. Burst retransmission and deflection routing were implemented and allowed only once. 42

43 TCP RTT vs. packet loss probability 43

44 value vs. packet loss probability 44

45 SAIMD Throughput over Barebone OBS 45

46 SAIMD Throughput over OBS with Burst Retransmission 46

47 SAIMD Throughput over OBS with Burst Deflection 47

48 SAIMD Fairness over OBS with Burst Retransmission 48

49 Analytical vs. Simulation Results 49

50 Delay-based TCP (Vegas) over OBS Networks Basem Shihada University of Waterloo

51 Background TCP Vegas Detects network congestion at earlier stages Reduces the sending rate linearly 20-50% fewer retransmissions Improves throughput 37% to 71% compared to TCP Reno OBS Burst contention occurs even at low traffic loads Bufferless One-way signaling Contention resolution schemes Burst deflection Burst retransmission Reduce burst loss probability Introduce additional delay 51

52 TCP Vegas Congestion Control 1. Compute the Expected throughput Expected (BaseRTT is the minimum measured RTT) 2. Compute the current Actual throughput cwnd Actual RTT 3. Compute the Diff = Expected Actual Diff cwnd(1 cwnd BaseRTT BaseRTT ) RTT 4. Adjust the next cwnd size as follows, cwnd 1 cwnd cwnd cwnd 1 diff diff. diff 52

53 TCP Vegas over OBS Vegas can NOT detect congestion in a barebone OBS network RTT in barebone OBS network varies little!!! What happens for Vegas over OBS network with burst retransmission or burst deflection? Case 1: low traffic loads Burst retransmission and burst deflection introduce higher burst delay Vegas detects sudden RTT increase due to retransmission or deflection Results in unnecessary reduction of the cwnd size! Case 2: high traffic loads Higher congestion in OBS results in even higher burst delay Vegas detects more frequent RTT increase 53

54 TCP Vegas Reengineering Goal: to design a proper TCP congestion avoidance mechanism that is suitable for OBS networks Observations: Congestion in access networks -> longer queuing delay -> continuous longer-rtts Congestion in OBS networks -> frequent deflections/retransmissions -> frequent longer-rtts No congestion -> less longer-rtts 54

55 Threshold-based TCP Vegas Threshold T is introduced to assist TCP Vegas to distinguish burst contentions at low traffic loads and network congestion Within N consecutive packets: minrtt(i) - the minimum RTT of the i consecutive packets (i<n) j - is a counter for recording the number of packets whose RTT > minrtt(i) T - is the threshold on the number of packets with longer RTT than the minrtt in the previous N rounds 55

56 Threshold-based TCP Vegas Flow Chart Reduce the sensitivity of TCP to RTT increase due to retransmission and deflection 56

57 Performance Modeling Goal: to understand the behavior of delay-based TCP over OBS networks Assumptions: High access bandwidth: TCP packets in a single cwnd are assembled to a single burst (i.e., TCP fast flow) IP access networks are not congested Burst loss probability and burst contention probability in OBS network are given Obtained: Vegas throughput over Barebone OBS OBS with burst retransmission 57

58 TCP Vegas over Barebone OBS Average RTT is very close to BaseRTT Slow-start-to-slow-start period (SS2SS): A to B (slow start), starts at 2 and doubles every other round B to C (transition period), increases by 1 every RTT due to Diff< C to D (loss free period), remains unchanged due to Diff TO period due to a burst loss 58

59 TCP Vegas over OBS with Burst Retransmission The cwnd that is retransmitted in the OBS layer has longer delay than BaseRTT SS2SS period: A to B (slow start) B to C (transition period) C to G (loss free period): the size of cwnd is modeled as Markov Chain TO period 59

60 TCP Vegas over OBS with Burst Retransmission (cont) The probability of a successful round that does not experience burst contention p The probability of a successful round in which a burst experiences contention but successfully retransmitted p nc sr 1 1 p p pc p 1 p c 60

61 Parameters T and N N and T should be much larger than the number of packets that are assembled into a single burst N = it, where i >1 if T is closer to N, less number of remaining packets in N to react to detected congestion state, which results in TCP ineffectively reacting to network congestion 61

62 Vegas Throughput over Barebone OBS 62

63 Vegas Throughput over OBS with Burst Retransmission 63

64 Simulation Results S 1 D 1 OBS Core Network D 2 S 2 C C 2 C 3 C 4 S x D x Multi-hop network topology Each link has 8 wavelengths, each of which operates at 10Gbps (OC-192) Burst assembly time is 5ms, burst-size threshold is 50KB, offset-time is 6µsec. TCP segment size is 1KB Burst retransmission is implemented. Bursts are retransmitted only once. T =100, 200, 300, 400 and N = 4T 64

65 Threshold-based Vegas over Barebone OBS 65

66 Threshold-based Vegas over OBS with Retransmission 66

67 Varying T and N with Burst Retransmission 67

68 Vegas Fairness 68

69 Analysis Notations 69

70 Analysis for TCP Vegas over Barebone OBS Number of packets sent from A to B Y AB W log 4 i 0 The duration from A to B Number of packets sent from B to C The duration from B to C 0 i logw logw0 A AB 2( ) R 2(logW0 2) R 4 Y BC W W0 3W i i W0 /2 W0 A BC ( ) R 2 70

71 Analysis for TCP Vegas over Barebone OBS Number of rounds in C to D S lossfree 1 p p 2(logW 1 p ( p Duration from C to D is 0, ACD RS lossfree 2(logW W 2) 2 otherwise. Number of packets sent from C to D is if YCD W0S lossfree Vegas TOP is similar to Reno Vegas throughput over barebone OBS is 0 0, 0 W0 2) ), 2 B barebone Y A AB AB Y A BC BC Y CD A CD E[ H ] TOP p E[ H] 1 p E TOP f ( p) RTO 1 p 71

72 Analysis for TCP Vegas over OBS with Burst Retransmission Average cwnd size from C to G E[ W CG ] W ( W W0 W i 0 [( W Duration from C to D is 1)... W i) ] Number of packets sent from C to D is 0 0 A ( p R p RTT ) S CG CG nc Y E[ W ] S 0 ' CG 0 0 sr i lossfree Vegas throughput over OBS with burst retransmission is 1 r ` 0 lossfree ( W 0 W 0 ') B ret Y A AB AB Y A BC BC Y CG A CG E[ H ] TOP 72

73 W 0 and W 0 W 0 -the cwnd size when Diff reaches from a smaller value W 0 the cwnd size when Diff reaches from a larger value 73