1 / 64

CIS 6930: DISTRIBUTED MULTIMEDIA SYSTEMS

CIS 6930: DISTRIBUTED MULTIMEDIA SYSTEMS. TCP SMOOTHNESS AND WINDOW ADJUSTMENT STRATEGY Authors: Chi Zhang and Vassilis Tsaoussidis Presented By: Mayank Srivastava Jinwook Choi. OUTLINE. Key terms used in the paper Introduction Discussion on TCP smoothness An experimental protocol

niles
Télécharger la présentation

CIS 6930: DISTRIBUTED MULTIMEDIA SYSTEMS

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CIS 6930: DISTRIBUTED MULTIMEDIA SYSTEMS TCP SMOOTHNESS AND WINDOW ADJUSTMENT STRATEGY Authors: Chi Zhang and VassilisTsaoussidis Presented By: MayankSrivastavaJinwookChoi

  2. OUTLINE Key terms used in the paper Introduction Discussion on TCP smoothness An experimental protocol Experiment and Analysis Conclusion

  3. KEY TERMS • AIMD - Additive increase/ Multiplicative Decrease • RAP- Rate adaptation protocol • RTT – Round trip time • α – Additive increase value • β – Multiplicative decrease ratio • TFRC – TCP Friendly Congestion Protocol

  4. OUTLINE Key terms used in the paper Introduction Discussion on TCP smoothness An experimental protocol Experiment and Analysis Conclusion

  5. INTRODUCTION-I • It is observed that when the system throughput is relatively stable, end users of media-streaming applications do not necessarily experience smooth throughput • This is due to the unsynchronized window adjustments triggered by random congestion indications. • In this paper the author: • Analyzes and evaluates the negative impact of random window adjustments on smoothness, short-term fairness, and long-term fairness.

  6. INTRODUCTION-I ( Contd.) 2. The author further proposes an experimental congestion avoidance mechanism, namely TCP(α, β, γ, δ), based on coordinated window adjustments. • The flow-level smoothness is enhanced significantly for media-streaming applications, without a cost on fairness and responsiveness. • Responsiveness is even boosted when bandwidth is underutilized.

  7. INTRODUCTION-II • Transmission control of standard TCP is based on AIMD window adjustment strategy. • Media-streaming applications find the standard multiplicative decrease by a factor of 2 upon congestion to be unnecessarily severe • In traditional AIMD, congestion window increases steadily until congestive packet loss is detected, which necessitates the subsequent error recovery. • In contrast, window adjustments of TCP Vegas take a congestion avoidance approach.

  8. INTRODUCTION-III • TCP friendly protocols favor smoothness for multimedia applications by using a gentle downward adjustment upon congestion, at the cost of lesser responsiveness through moderated upward adjustments. • RAP decouples congestion control from application reliability, smoothness is not a design criterion. • TCP friendly rate control (TFRC) is an equation-based TCP-friendly congestion control protocol for unicast applications.

  9. INTRODUCTION III (Contd.) • GAIMD is a TCP-friendly protocol that generalizes AIMD congestion control by parameterizing the additive increase value and multiplicative decrease ratio . • It derives a simple relationship between α and β to be friendly to standard TCP, α=1, β=1/2. • The challenge does not lie in simply achieving smooth transmission control, but rather in providing smoothness along with bandwidth efficiency, fairness, responsiveness, as well as controlled queue lengths.

  10. INTRODUCTION IV • The measurement-based adaptive congestion avoidance mechanism , they use is TCP(α, β, γ , δ), to improve flow-level throughput smoothness. • The mechanism coordinates the upward and downward window adjustments to abolish the damage of unsynchronized and random window control on throughput smoothness. • While the smoothness of TCP Vegas is achieved at the expense of fairness, the new mechanism enhances significantly both smoothness and fairness, without compromising responsiveness.

  11. OUTLINE Key terms used in the paper Introduction Discussion on TCP smoothness An experimental protocol Experiment and Analysis Conclusion

  12. DYNAMICS OF CONGESTION CONTROL • Consider a simple network topology shown in the figure, in which the link bandwidth and propagation delay are labeled.

  13. ‘n’ TCP flows share a bottleneck link with capacity of ‘bw’. • RTT₀ = 2 *(src_delay +delay + sink_delay) • For the overall system behavior, they define the aggregated congestion window of the system at time as:-

  14. The system throughput is given as:- where qdelay(t)is the queuing delay at the bottleneck router R1. • Thus throughput is not only a function of the congestion window, but also a function of the dynamic queuing delay

  15. Assuming all flows are in the additive increase stage, if scwnd(t) is below the point knee: • There is no queue buildup (i.e. RTT(t)= RTT0 ), but there could be temporary queue buildup due to the traffic burstiness. • The throughput grows in proportion to scwnd ,since the bottleneck capacity is not fully utilized until scwnd increases to scwndknee

  16. If scwnd (t) increases further beyond scwndknee, however, the system displays different dynamics. • The bottleneck queue starts to build up, after the bottleneck capacity is saturated. • scwnd(t) can be represented like below: • Steady queuing delay will be:

  17. The system throughput is bounded by the physical capacity bw , in spite of the increase of scwnd(t) beyond the knee, because qdelay(t) in the denominator of (2) grows as well. • The system dynamics can be continuously described by (4)–(6),until the queue length ∆w(t) reaches the max buffer size i.e., when scwnd touches the point cliff.

  18. Equation (6) shows that increasing scwnd beyond scwndkneedoes not further enhance system throughput, but only results in high queuing delay. • Multiplicative decrease is necessary to accomplish fairness dynamically, but it does not sacrifice the system throughput, as long as scwnd operates between scwndkneeand scwndcliff. • To prevent scwnd from operating below scwndknee where bandwidth is underutilized, and meanwhile maintain adequate AIMD oscillation, an efficient window decreasing ratio would be:

  19. Here scwnd will decrease from the cliff down to the knee (for sufficient fairness-oriented AIMD oscillation), but not below the knee. • Equation (8) implies that an efficient decrease ratio depends on the network settings.

  20. OBSERVATIONS ON IMPROVING TCP SMOOTHNESS • ‘scwnd’ operating between the knee and the cliff where the overall system throughput is maximum and relatively stable does not imply that each flow will observe a smooth throughput. • The authors are particularly interested in the impact of random window adjustments on the flow smoothness of each flow • The throughput of the ithflow at time t is:-

  21. Unsynchronized multiplicative decrease degrades the short-term fairness (measured in several RTTs), due to random congestive drops. • Flows consuming extra bandwidth at one time period must pay back the credit to the flows consuming less bandwidth, at some other time period. • As a result of the long-term fairness accomplished without short-term guarantees, individual flows unavoidably see throughput oscillations, even when the system throughput is stable.

  22. Hence , in this paper, the authors observe that the major obstacle for achieving smoothness with individual flows is unsynchronized and random window adjustments. • If upward and downward window adjustments are synchronized in equilibrium, however, short-term fairness is not damaged. • With a stable system throughput, the bandwidth allocated to each end user will be also smooth over time. • Therefore, smoothness can be achieved along with bandwidth efficiency and fairness.

  23. OUTLINE Key terms used in the paper Introduction Discussion on TCP smoothness An experimental protocol Experiment and Analysis Conclusion

  24. AN EXPERIMENTAL PROTOCOL • Based on previous slides, the author proposes a novel congestion avoidance mechanism to improve the TCP smoothness. • The sender records the minimum RTT and the maximum RTT perceived. • The queuing delay can be derived by deducting min RTT(which corresponds to the roundtrip propagation delay) from the current RTT.

  25. In the congestion avoidance stage, the additive increase speed (α=1) of standard TCP is untouched, and the sender halves the congestion window upon packet losses (β=0.5). • The standard congestion control is complemented with the following congestion avoidance mechanism upon the detection of the following condition:

  26. This shows that multiplicative decrease is triggered when scwnd is halfway between scwndkneeand scwndcliffand swnd decreases to scwndknee.

  27. Another optional control parameter δ is introduced, in order to enhance the additive increase speed when scwnd operates below the knee. • At the above condition, the queue is relatively close to empty and very likely the bandwidth is underutilized. • Thupperprovides an upper bound on the queue length and Thlowerdefines whether the queue is to close to idle and should be used.

  28. One concern with RTT-based congestion avoidance is the accuracy of min RTT. • min RTT overestimation is a well-known problem for TCP Vegas, as it stabilizes at a nonempty queue. • With proposed protocol, however, scwnd periodically adjusts downward to scwndkneewhere RTT0 be accurately measured. • Now in the following slides, the system convergence is analyzed theoretically.

  29. Assume when the ith flow joins the system, its min RTT is overestimated by ∆RTT , due to a bottleneck queue with q0 = ∆RTT * bw bytes. • According to (11) [26], the next γ -based window decrease will be triggered when scwnd(t)= scwndknee/γ. • After that, the expected in the next RTT will be:-

  30. where (according to (11)) • The window decrease ratio γifor the ithflow is higher than the γ for other flows due to the overestimation of min RTT. • The queue size right after the coordinated window decreases will be:

  31. (15) implies that a flow overestimating initially will quickly converge to an accurate estimation.

  32. SUMMARY OF THIS SECTION • TCP(α, β, γ , δ) has a number of features:- • Relies on fine-grained RTT measurements to estimate network conditions, beyond congestive packet drops. • Follows a synchronized model by coordinating upward and downward window adjustments to enhance TCP smoothness, without compromising efficiency, fairness and responsiveness.

  33. Avoids congestion by scheduling window decreases well before the occurrence of congestion, to reduce packet drops. • Introduces new control parameters γ and δ and adaptable to the current network condition. • It is an end-to-end solution that controls the average queue length through Thupper and Thlower without deploying AQM (active queue management) in routers.

  34. EVALUATION AND ANALYSIS • TESTING PLAN AND METHODOLOGY

  35. In the figure, the values by default are assumed to be : delay=30ms and src_delay=sink_delay=5ms and bw varies from 10Mbps to 100Mbps. • Protocol behaviors are tested with multiple bottlenecks and cross traffic shown in Fig. 2. • To evaluate the protocol’s responsiveness to the rapid changes of available bandwidth, the authors created a scenario of temporary “blackouts” due to mobile handoffs, during which all transmitted packets are lost. • An error model is inserted into the access links to the receivers (assumed in the wireless domain), with 100% packet dropping rate.

  36. Seven protocol configurations are selected: Reno (α=1 ,β=0.5), Reno with RED, TCP Vegas, Reno plus γ , Reno plus γ andδ(=2) and , smooth TCP (α=0.31 ,β=0.875 ) , and TFRC • PERFORMANCE METRICS • Long-term fairness is measured by the: • For a “worst-case” analysis and a tighter bound on fairness, the authors propose to use:-

  37. The author uses Coefficient of Variation ( CoV) to gauge the throughput smoothness experienced by flow : • Allotted throughput is also used to compute the short-term fairness, derived from the traditional Fairness Index:

  38. Allotted throughput, and hence CoV and short-term fairness, are measured 15 s after the simulation starts, in order to just capture the system behaviors after it enters equilibrium.

  39. RESULTS AND ANALYSIS FAIRNESS AND SMOOTHNESS The authors first conduct ten simulations for 100–s connection time, with the number of flows varied from 10 to 100.

  40. Random multiplicative decrease can not guarantee worst-case fairness in the above figure, compared to. the synchronized window adjustments of Reno+γ

  41. 2.Diverse RTTs and Multiple Bottlenecks: • The propagation delays of access links to sink nodes are uniformly distributed between 5 ms to 15 ms. The buffer size is set to the product of the bottleneck capacity and the minimum round-trip propagation delay. • The worst-case fairness and CoV are demonstrated in the following figures:-

  42. Protocols are further tested with multiple bottlenecks and cross traffic. Half of the flows form the main traffic, while the other half form the cross traffic.

  43. 3) Responsiveness With Dynamic Traffic: • Our next simulation over heterogeneous (wires/wireless) networks indicates that the high smoothness of coordinated window adjustments is achieved not at the cost of responsiveness. • A single flow traverses a 10 Mbps bottleneck link. At time 20 s, the wireless access link is interrupted by a 1 second handoff period, during which all packets are lost. • Since bandwidth becomes available immediately after the handoff, a high sending rate increase is the desired behavior.

  44. The protocols’ aggressiveness with allotted throughputs after the handoff is shown in the above figure.

  45. 4) Microscopic Behaviors: • Microscopic behaviors in a static environment: • The author tracked the allotted throughputs of three flows competing at a 3 Mbps bottleneck link for 100 s, in order to observe the microscopic behaviors of protocols. • The number of flows (3) is selected to ease the analysis. The allotted throughput of each flow is traced as well as the aggregated system throughput, with a 200 ms sampling period.

More Related