## Packet-Pair Flow Control (1994)

Venue: | IEEE/ACM Transactions on Networking |

Citations: | 44 - 2 self |

### BibTeX

@ARTICLE{Keshav94packet-pairflow,

author = {Srinivasan Keshav},

title = {Packet-Pair Flow Control},

journal = {IEEE/ACM Transactions on Networking},

year = {1994}

}

### OpenURL

### Abstract

This paper presents the packet-pair rate-based feedback flow control scheme. This scheme is designed for networks where individual connections do not reserve bandwidth and for the available bitrate (best-effort) component of integrated networks. We assume a round-robin-like queue service discipline in the output queues of the network's switches, and propose a linear stochastic model for a single conversation in a network of such switches. These model motivates the Packet-Pair rate probing technique, which forms the basis for provably stable discrete and continuous time rate-based flow control schemes. We present a novel state estimation scheme based on fuzzy logic. We then address several practical concerns: dealing with system startup, retransmission and timeout strategy, and dynamic setpoint probing. We present a finite state machine as well as source code for a model implementation. The dynamics of a single source, the interactions of multiple sources, and the behavior of packet-pai...

### Citations

3058 | Fuzzy sets
- Zadeh
- 1965
(Show Context)
Citation Context ...ck the changes in the system. To summarize, the observation is that if the predictor error is large, then a should be small, and vice versa. Treating `small' and `large' as fuzzy linguistic variables =-=[62]-=-, this is the basis for a fuzzy controller for the estimation of a. The controller implements three fuzzy laws: If proportional error is low, then a is high If proportional error is medium, then a is ... |

2265 |
Congestion avoidance and control
- Jacobson
- 1988
(Show Context)
Citation Context ...ting window is 200 packets, this scheme will take 200 round trip times to reach that size. A faster way to reach the operating point is the exponential + linear scheme proposed by Jacobson and Karels =-=[26]-=-. However, the choice of when to change phase from exponential to linear increase is rather ad hoc. Finally, Kanakia and Mishra have proposed a linear increase in the sending rate till the operating p... |

1672 | A generalized processor sharing approach to flow control in integrated services networks: The single node case
- Parekh, Gallager
- 1993
(Show Context)
Citation Context ...hown that there is a strong motivation to use round-robin-like packet schedulers for bursty data traffic, since this provides several advantages over the traditional first-come-firstserved discipline =-=[12, 24, 34, 44, 46]-=-. Round-robin service automatically enforces a min-max fair allocation of resources [20]. It also automatically polices sources, since a source sending faster than its fair share is the one that will ... |

1078 | Analysis and simulation of a fair queueing algorithm
- Demers, Keshav, et al.
- 1990
(Show Context)
Citation Context ...s equally applicable to existing datagram-oriented networks. Congestion control for ABR traffic involves packet (or cell) schedulers at queuing points and the flow control protocol at traffic sources =-=[12, 34]-=-. We first consider the role of packet schedulers in congestion control. Recent work has shown that there is a strong motivation to use round-robin-like packet schedulers for bursty data traffic, sinc... |

551 |
Outline of a new approach to the analysis of complex systems and decision processes
- Zadeh
- 1973
(Show Context)
Citation Context ...n. It would be useful to automatically determine a `good' value of a, and to be able to change this value on-line if the system behavior changes. Our approach uses fuzzy control to effect this tuning =-=[63, 67]-=-. Fuzzy exponential averaging uses the assumption that a system can be thought of as belonging to a spectrum of behavior that ranges from `steady' to `noisy'. In a `steady' system, the sequence {q k }... |

382 | A control-theoretic approach to flow control
- Keshav
- 1991
(Show Context)
Citation Context ...ys. 4.3.1. Fuzzy Prediction Having derived the control law, and proved its stability, we now need to determine stable estimators for the system state. In earlier work, we studied the Kalman estimator =-=[32]-=-. However, this estimator requires a priori knowledge of the variances of the system and observation noise. Since these values are hard to determine, we turn to a novel technique for state estimation,... |

340 | Some observations on the dynamics of a congestion control algorithm
- Schenker, Zhang, et al.
- 1990
(Show Context)
Citation Context ... Two-Way Traffic A traffic scenario with traffic flowing in both directions through a switch is important to examine because of the role played by acknowledgement packets in some flow control schemes =-=[66]-=-. The TCP scheme uses acks as a clocking mechanism to inject packets into the network at roughly the bottleneck rate. The packet-pair scheme uses the gap between acknowledgment packets to estimate the... |

336 |
Sondhi: Stochastic Theory of a Data-Handling System with Multiple Sources
- Anick, Mitra, et al.
(Show Context)
Citation Context ...rk in continuous time. We also make the fluid approximation [1], so packet boundaries are ignored, and the data flow is like that of a fluid in a hydraulic system. This approximation is commonly used =-=[3, 6, 40, 55]-=-, and both analysis [43] and simulations show that the approximation is a close one, particularly when the bandwidth-delay product is large. Let us assume that the input rate l is held fixed for some ... |

312 | A Binary Feedback Scheme for Congestion Avoidance in Computer Networks with a Connectionless Nctwork Layer
- Rarnaknshnan, Jain
- 1988
(Show Context)
Citation Context ...y product, having a good startup scheme is crucial. Several startup schemes have been described in the literature. The DECbit scheme is to open the window linearly till the operating point is reached =-=[48]-=-. This is a poor choice when the bandwidth delay product is large. For example, when the operating window is 200 packets, this scheme will take 200 round trip times to reach that size. A faster way to... |

186 |
An adaptive congestion control scheme for real-time packet video transport
- Kanakia, Mishra, et al.
- 1993
(Show Context)
Citation Context ...use in public data networks. 9.5. Packet Video Using Packet-pair In the final test of packet-pair, we used it to carry MPEG compressed video over a simulated network. The experimental method is as in =-=[30]-=-, where software coded MPEG is carried through a simulated network with 8 active sources, limited buffering and sudden changes in available bandwidth. The feedback signals from the flow control algori... |

122 |
A New Architecture for Packet Switching Network Protocols
- Zhang
- 1989
(Show Context)
Citation Context ...s fixed (as in ATM networks), or service is weighted to reflect the packet sizes- (as with Fair Queueing [12], Weighted Fair Queueing, Packetized Generalized Processor Sharing [46], and Virtual Clock =-=[65]-=-). We call this kind of round-robin server a Rate-Allocating Server, (RAS) since at each round a connection is allocated a service rate depending on the number of active connections and their weights,... |

88 |
NETBLT: A Bulk Data Transfer Protocol," RFC 998, Network Working Group
- Clark, Lambert, et al.
- 1987
(Show Context)
Citation Context ...he performance of generic dynamic window schemes. 12.2. Rate-based Flow Control Several rate-based flow control schemes have been proposed in the literature. One of the earliest scheme was for NETBLT =-=[11]-=-, which was based on heuristics, and not thoroughly studied. In NETBLT, the transmitter sends data at some rate for a few round trip times, and the receiver clocks the data to see if it was received a... |

85 |
Congestion Control in Computer Networks
- Keshav
- 1991
(Show Context)
Citation Context ...s equally applicable to existing datagram-oriented networks. Congestion control for ABR traffic involves packet (or cell) schedulers at queuing points and the flow control protocol at traffic sources =-=[12, 34]-=-. We first consider the role of packet schedulers in congestion control. Recent work has shown that there is a strong motivation to use round-robin-like packet schedulers for bursty data traffic, sinc... |

61 |
A Theoretical Analysis of Feedback Flow Control
- Shenker
- 1990
(Show Context)
Citation Context ... dynamics of a system where users update their sending rate either synchronously or asynchronously in response to measured round trip delays, or explicit congestion signals, for example in References =-=[5, 7, 8, 13, 52]-=-. These approaches typically assume Poisson sources, availability of global information, a simple flow update rule, and exponential servers. We do not make such assumptions. Further, they deal with th... |

50 |
Fuzzy Set Theory –and its Applications
- Zimmerman
- 1991
(Show Context)
Citation Context ...n. It would be useful to automatically determine a `good' value of a, and to be able to change this value on-line if the system behavior changes. Our approach uses fuzzy control to effect this tuning =-=[63, 67]-=-. Fuzzy exponential averaging uses the assumption that a system can be thought of as belonging to a spectrum of behavior that ranges from `steady' to `noisy'. In a `steady' system, the sequence {q k }... |

47 |
TCP timers don’t work well
- ZHANG
- 1986
(Show Context)
Citation Context ...er, if timers are too small there can be numerous retransmissions, which can lead to network congestion. A poor choice in the other direction will lead to long pauses, wasting the available bandwidth =-=[64]-=-. Thus, the choice of timers and of which packets to retransmit is intimately related to flow and congestion control. Three considerations guided the design of the timeout and retransmission strategy ... |

44 |
Dynamic adaptive windows for high speed data networks: theory and simulations
- Mitra, Seery
- 1990
(Show Context)
Citation Context ...the setpoint can deal with the case where B is unknown. Work by Mitra et al has shown that asymptotic analysis of product form queueing networks can be used to derive an optimal value of the setpoint =-=[42, 43]-=-. While their ideas are not directly applicable because of their assumptions of FCFS scheduling, Poisson cross traffic and an exponentially distributed packet service time distribution, to a first app... |

40 |
Modeling and Control of Dynamic Flows in Communication Networks
- Filipiak
- 1988
(Show Context)
Citation Context ... RAS service isolates a conversation's service from the details of the arrival pattern of cross traffic. For FCFS service, there is no such isolation, and the associated non-linear control is complex =-=[17, 55]-=-.s- 3. Packet-pair probing SOURCE SERVER 1 BOTTLENECK SINK RTT s rate estimate d e i j l m bottleneck rate Figure 1: Packet-pair probing The packet-pair mechanism is to send all data in the form of ba... |

36 | Dynamic Control of Session Input Rates in Communication Networks
- Gafni, Bertsekas
- 1984
(Show Context)
Citation Context ..., since this provides several advantages over the traditional first-come-firstserved discipline [12, 24, 34, 44, 46]. Round-robin service automatically enforces a min-max fair allocation of resources =-=[20]-=-. It also automatically polices sources, since a source sending faster than its fair share is the one that will be subjected to packet or cell loss. This ensures that well-behaved users are protected ... |

36 | Strikwerda, “Analysis of dynamic congestion control protocols—A Fokker–Planck approximation
- Mukherjee, C
(Show Context)
Citation Context ...ays lead their control to oscillate even under stable conditions. Dynamic window schemes have been mathematically analyzed by Bolot [6], Fendick, Rodrigues and Weiss [15] and Mukherjee and Strikwerda =-=[45]-=-. These analyses give some insight into the performance of generic dynamic window schemes. 12.2. Rate-based Flow Control Several rate-based flow control schemes have been proposed in the literature. O... |

34 |
Alan weiss“Analysis of rate-based control strategy for long haul data transport”, Performance Evaluation 16
- Fendick, Rodrigues
- 1992
(Show Context)
Citation Context ...ictive control, so propagation delays lead their control to oscillate even under stable conditions. Dynamic window schemes have been mathematically analyzed by Bolot [6], Fendick, Rodrigues and Weiss =-=[15]-=- and Mukherjee and Strikwerda [45]. These analyses give some insight into the performance of generic dynamic window schemes. 12.2. Rate-based Flow Control Several rate-based flow control schemes have ... |

34 |
Numerical methods for modeling computer networks under nonstationary conditions
- Tipper, Sundareshan
- 1990
(Show Context)
Citation Context ... RAS service isolates a conversation's service from the details of the arrival pattern of cross traffic. For FCFS service, there is no such isolation, and the associated non-linear control is complex =-=[17, 55]-=-.s- 3. Packet-pair probing SOURCE SERVER 1 BOTTLENECK SINK RTT s rate estimate d e i j l m bottleneck rate Figure 1: Packet-pair probing The packet-pair mechanism is to send all data in the form of ba... |

30 |
Rate Controlled Servers for Very High Speed Networks
- Kalmanek, Kanakia, et al.
- 1990
(Show Context)
Citation Context ...eing kept idle if necessary (i.e. a non-work-conserving discipline), this is called framed round robin service. A multilevel framed round robin server is also called a hierarchical round robin server =-=[28]-=-. The network model (and hence the flow control scheme) described in this paper is adequate to describe variants of round-robin service where the packet size per connection is fixed (as in ATM network... |

28 |
How fair is fair queueing
- Greenberg, Madras
- 1990
(Show Context)
Citation Context ...nd then serving packets in increasing order of their tags, as in Fair Queueing [12]. It can be shown that Fair Queueing emulates head-of-line processor sharing asymptotically with conversation length =-=[21, 22]-=-. Connections may be allocated unequal bandwidth shares by varying the weights, as is done in the Weighted Fair Queueing discipline. If the round spans a constant time duration, with the output trunk ... |

28 |
A simple theory of traffic and resource allocation
- Low, Varaiya
- 1991
(Show Context)
Citation Context ...rk in continuous time. We also make the fluid approximation [1], so packet boundaries are ignored, and the data flow is like that of a fluid in a hydraulic system. This approximation is commonly used =-=[3, 6, 40, 55]-=-, and both analysis [43] and simulations show that the approximation is a close one, particularly when the bandwidth-delay product is large. Let us assume that the input rate l is held fixed for some ... |

26 | Analysis of a fluid approximation to flow control dynamics
- Bolot, Shankar
- 1992
(Show Context)
Citation Context ...rk in continuous time. We also make the fluid approximation [1], so packet boundaries are ignored, and the data flow is like that of a fluid in a hydraulic system. This approximation is commonly used =-=[3, 6, 40, 55]-=-, and both analysis [43] and simulations show that the approximation is a close one, particularly when the bandwidth-delay product is large. Let us assume that the input rate l is held fixed for some ... |

24 |
Asymptotically optimal design of congestion control for high speed data network
- Mitra
- 1992
(Show Context)
Citation Context ...the setpoint can deal with the case where B is unknown. Work by Mitra et al has shown that asymptotic analysis of product form queueing networks can be used to derive an optimal value of the setpoint =-=[42, 43]-=-. While their ideas are not directly applicable because of their assumptions of FCFS scheduling, Poisson cross traffic and an exponentially distributed packet service time distribution, to a first app... |

21 |
Queueing Disciplines and Passive Congestion Control in Byte-Stream Networks
- Morgan
- 1989
(Show Context)
Citation Context ...hown that there is a strong motivation to use round-robin-like packet schedulers for bursty data traffic, since this provides several advantages over the traditional first-come-firstserved discipline =-=[12, 24, 34, 44, 46]-=-. Round-robin service automatically enforces a min-max fair allocation of resources [20]. It also automatically polices sources, since a source sending faster than its fair share is the one that will ... |

17 |
Feedback control of congestion in store-and-forward networks: the case of a single congested node
- Benmohamed, Meerkov
- 1993
(Show Context)
Citation Context ...the phase 4 buildup is smaller since the rate probes are more likely to interfere with competing sources than when N = 2. Thus, the peak phase 4 buildup is achieved when N = 2. To summarize, for N in =-=[2,4]-=-, there are no losses, and the peak is for N = 2, when there is least interference. For N in [5,7], there are packet losses in phase 2, as well as queue draining in phase 3, and the subsequent phase 4... |

17 |
Loss-load curves: support for rate-based congestion control in high-speed datagram networks
- Williamson, Cheriton
- 1991
(Show Context)
Citation Context ...sing a fuzzy predictor. Further, they appeal primarily to intuitive heuristics, and do not use a formal control-theoretic model, to develop their control. Williamson has proposed the Loss-Load scheme =-=[60, 61]-=- that uses the throughput-versus-loss curve to compute an optimal sending rate. This approach has numerous lacunae. It ignores system considerations, such as the the fact that monitoring each connecti... |

16 |
Dynamic Modeling and Control of Congestion-prone Systems
- Agnew
- 1976
(Show Context)
Citation Context ...ckly as once every packet-pair with no change to the length of the state vector.s- When control is done once every probe, it is easier to work in continuous time. We also make the fluid approximation =-=[1]-=-, so packet boundaries are ignored, and the data flow is like that of a fluid in a hydraulic system. This approximation is commonly used [3, 6, 40, 55], and both analysis [43] and simulations show tha... |

16 | Fuzzy Prediction of Time Series
- Khedkar, Keshav
- 1992
(Show Context)
Citation Context ...variance in the observation noise is small can now be removed. The resulting system is shown in Figure 4. Further details of the prediction system and a performance analysis can be found in Reference =-=[37]-=-. Exponential Averager Exponential Averager Fuzzy System Fuzzy System b z - 1 z - 1 z - 3 a Observation Proportional error Smoothed proportional error Estimate qsq Figure 4: Fuzzy prediction system 5.... |

15 |
Xunet 2: A Nationwide Testbed
- Fraser, Kalmanek, et al.
- 1992
(Show Context)
Citation Context ...en all the queueing points are round-robin like, and there is no multiplexing of data streams in the network. There are high-speed ATM networks already in existence that implement round-robin service =-=[19, 51]-=-. Further, the proposed native-mode ATM stack does not do any multiplexing so as to preserve per-virtual circuit Quality of Service constraints. Thus, we feel that these two restrictions could easily ... |

15 | A Scheduling Discipline and Admission Control Policy for Xunet 2
- Saran, Keshav, et al.
(Show Context)
Citation Context ...en all the queueing points are round-robin like, and there is no multiplexing of data streams in the network. There are high-speed ATM networks already in existence that implement round-robin service =-=[19, 51]-=-. Further, the proposed native-mode ATM stack does not do any multiplexing so as to preserve per-virtual circuit Quality of Service constraints. Thus, we feel that these two restrictions could easily ... |

14 |
A New Approach to Performance-Oriented Flow Control
- Bharath-Kumar, Jaffe
- 1981
(Show Context)
Citation Context ...ources than when N = 2. Thus, the peak phase 4 buildup is achieved when N = 2. To summarize, for N in [2,4], there are no losses, and the peak is for N = 2, when there is least interference. For N in =-=[5,7]-=-, there are packet losses in phase 2, as well as queue draining in phase 3, and the subsequent phase 4 buildup peaks when N = 7. Finally, the phase 2 buildup peaks when N = 10. _______________________... |

14 |
Decentralized algorithms for optimal flow control
- Bovopoulos, Lazar
- 1987
(Show Context)
Citation Context ...ources than when N = 2. Thus, the peak phase 4 buildup is achieved when N = 2. To summarize, for N in [2,4], there are no losses, and the peak is for N = 2, when there is least interference. For N in =-=[5,7]-=-, there are packet losses in phase 2, as well as queue draining in phase 3, and the subsequent phase 4 buildup peaks when N = 7. Finally, the phase 2 buildup peaks when N = 10. _______________________... |

13 |
Bandwidth Management: A Congestion Control Strategy for Broadband Packet Networks
- Eckberg, Luan, et al.
- 1989
(Show Context)
Citation Context ...em, and for adequate responsiveness, it should not be too small. Simulations indicate that a value of 0.9 is a good compromise between responsiveness and instability. Similar studies are mentioned in =-=[14]-=-. 4.3. Controller Design: Continuous Time Control This section describes how the frequency of control can be increased by using information about the propagation delay. Note that nsb , the estimate fo... |

13 |
Design and analysis of a flow control algorithm for a network of rate allocating servers
- Keshav, Agrawala
- 1991
(Show Context)
Citation Context ...ects the lack of information about the rate change for one round trip time. The amount of buildup is exactly the minimum predicted by multiplying the change in the service rate by the round trip time =-=[33]-=-. However, phase 4 buildup because of improper rate probing when several sources are in their startup phase, was somewhat unexpected. As it turns out, this effect dominates peak buffer occupancy. Four... |

12 |
On Pareto-optimal flow control in an integrated environment
- Douligeris, Mazumdar
- 1986
(Show Context)
Citation Context ... dynamics of a system where users update their sending rate either synchronously or asynchronously in response to measured round trip delays, or explicit congestion signals, for example in References =-=[5, 7, 8, 13, 52]-=-. These approaches typically assume Poisson sources, availability of global information, a simple flow update rule, and exponential servers. We do not make such assumptions. Further, they deal with th... |

12 |
Optimal flow control of multi-class queueing networks with decentralized information
- Hsiao, Lazar
- 1987
(Show Context)
Citation Context ...[57] have considered the the design of optimal traffic filters when the state is not fully observable, but the filters are specialized for voice traffic. Robertazzi and Lazar [49] and Hsiao and Lazar =-=[25]-=- have shown that under a variety of conditions, the optimal flow control for a Jacksonian network with Poisson traffic is bang-bang (approximated by a window scheme). It is not clear that this result ... |

12 |
Optimizing File Transfer Response Time Using Loss-Load Curves: A Simulation Study
- Williamson
- 1995
(Show Context)
Citation Context ...sing a fuzzy predictor. Further, they appeal primarily to intuitive heuristics, and do not use a formal control-theoretic model, to develop their control. Williamson has proposed the Loss-Load scheme =-=[60, 61]-=- that uses the throughput-versus-loss curve to compute an optimal sending rate. This approach has numerous lacunae. It ignores system considerations, such as the the fact that monitoring each connecti... |

11 |
Fairness and congestion control on a large ATM data network with dynamically adjustable windows
- Hahne, Kalmanek, et al.
- 1991
(Show Context)
Citation Context ...idth-delay product, and hence the required window can be large (of the order of hundreds of kilobytes per conversation). In view of this, the adaptive window allocation scheme proposed by Hahne et al =-=[23]-=- is attractive. In their scheme, a conversation is allocated a flow control window that is always larger than the product of the allocated bandwidth at the bottleneck, and the round trip propagation d... |

11 |
Robin Scheduling for Fair Flow Control
- Hahne, Gallager, et al.
- 1986
(Show Context)
Citation Context ...hown that there is a strong motivation to use round-robin-like packet schedulers for bursty data traffic, since this provides several advantages over the traditional first-come-firstserved discipline =-=[12, 24, 34, 44, 46]-=-. Round-robin service automatically enforces a min-max fair allocation of resources [20]. It also automatically polices sources, since a source sending faster than its fair share is the one that will ... |

9 | Warp control: a dynamically stable congestion protocol and its analysis
- Park
- 1993
(Show Context)
Citation Context ...ilar to that we derive in Section 4, thus validating some of our work. However, he has not done detailed simulations to confirm his control law. Park has proposed the 'Warp' congestion control scheme =-=[47]-=-. In this scheme, the receiver detects network state by observing arrival times of data packets from the sender, and uses this to determine an appropriate sending rate. However, this work suffers from... |

9 |
A high speed transport protocol for datagram /virtual circuit networks
- Sabnani, Netravali
- 1989
(Show Context)
Citation Context ...This has the effect of increasing the estimate of . Note that this source of error will persist even if the inter-ack spacing is noted at the sink and sent to the source using a state exchange scheme =-=[50]-=-. Measuringsb at the sink will reduce the effect of noise, but cannot eliminate it, since any server that is after the bottleneck potentially introduces noise in the measurement. We model this error i... |

9 |
Flow control protocols for integrated networks with partially observed voice traffic
- Vakil, Lazar
- 1987
(Show Context)
Citation Context ... and though they propose the use of recursive minimum mean squared error filters to estimate system state, the bulk of the results assume complete information about the network state. Vakil and Lazar =-=[57]-=- have considered the the design of optimal traffic filters when the state is not fully observable, but the filters are specialized for voice traffic. Robertazzi and Lazar [49] and Hsiao and Lazar [25]... |

8 |
Designing a public data network
- Fraser
- 1991
(Show Context)
Citation Context ...ter than its fair share is the one that will be subjected to packet or cell loss. This ensures that well-behaved users are protected from ill-behaved users, which is desirable in public data networks =-=[18]-=-. Due to these advantages, it is likely that round-robin schedulers will be widely implemented. packet-pair flow control is one way to exploit the properties of such schedulers to do intelligent flow ... |

8 | Myths About Congestion Management - Jain - 1992 |

8 | Semantics and implementation of a native-mode ATM protocol stack
- Keshav, Saran
- 1995
(Show Context)
Citation Context ...f the sender. 10.1. Packet-pair and ATM The packet-pair scheme is designed to work well in ATM environments. We propose to place packet-pair at the transport layer of a native-mode ATM protocol stack =-=[36]-=-. User data would be handed by the session layer to the transport layer using the transmission queue described in Section 6. The transport layer would then use packet-pair to send out pairs of AAL5 fr... |

6 |
Comparison of a fair queueing discipline to processor sharing
- Greenberg, Madras
- 1990
(Show Context)
Citation Context ...nd then serving packets in increasing order of their tags, as in Fair Queueing [12]. It can be shown that Fair Queueing emulates head-of-line processor sharing asymptotically with conversation length =-=[21, 22]-=-. Connections may be allocated unequal bandwidth shares by varying the weights, as is done in the Weighted Fair Queueing discipline. If the round spans a constant time duration, with the output trunk ... |