Citation Context

...trol Protocol (TCP) in source and destination computers involved in data transfers. The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 =-=[20]-=- and has gone through several enhancements since, e.g., [21], [58], [47], [18], [1], [14]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Interne...

Citation Context

...prove that FAST TCP has the same equilibrium properties as TCP Vegas [50], [44]. In particular, it does not penalize flows with large propagation delays and it achieves weighted proportional fairness =-=[31]-=-. For the special case of single bottleneck link with heterogeneous flows, we prove that the window control algorithm of FAST is locally asymptotically stable, in the absence of feedback delay. In Sec...

Citation Context

..., since the utility functions are strictly concave and is full rank (see, e.g., [42]). We need to show that the dual problem of (6) is indeed given by (7). Now the dual objective function is given by =-=[2]-=- Since the last term is independent of , minimizing over is the same as minimizing (7) over . Hence, there exists a unique solution for (6)–(7). 8 cf. the proof of a similar result in [50].s1256 IEEE/...

Citation Context

...urements. 1) Throughput: The average aggregate throughput for the interval [1, m] is defined as E := �n i=1 xi. 2) Intra-protocol fairness: Jain’s fairness index for the interval [1, m] is defined as =-=[22]-=- F := ( �n i=1 xi) 2 /(n �n i=1 x2i ). F ∈ (0, 1] and F = 1 is ideal (equal sharing). 7s3) Stability: The stability index of flow i is the sample standard deviation normalized by the average throughpu...

Citation Context

...products [19], [39]. These must be addressed by different means. Congestion window can be stabilized only if multi-bit feedback is used. 3 This is the approach taken by the equationbased algorithm in =-=[13]-=-, where congestion window is adjusted based on the estimated loss probability in an attempt to stabilize around a target value given by (2). This approach eliminates the oscillation due to packet-leve...

Citation Context

...lem of Reno, the window control algorithm must be stable, in addition to being fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [31], =-=[42]-=-, [35], [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation...

Citation Context

... both analytically and experimentally. In Section III-B, we present a mathematical model of the window control algorithm. We prove that FAST TCP has the same equilibrium properties as TCP Vegas [44], =-=[50]-=-. In particular, it does not penalize flows with large propagation delays and it achieves weighted proportional fairness [31]. For the special case of single bottleneck link with heterogeneous flows, ...

Citation Context

...ty, in period . 2 Here, is the peak window size that gives the “average” window of . Hence, a flow-level model of AIMD is Setting in (1) yields the well-known formula for TCP Reno discovered in [37], =-=[48]-=-, which relates loss probability to window size in equilibrium: 1 It should be @I 0 � @�AA packets, where � @�A is the end-to-end loss probability. This is roughly 1 when � @�A is small. 2 This model ...

Citation Context

...ongestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 [20] and has gone through several enhancements since, e.g., [21], [58], [47], [18], [1], =-=[14]-=-. It has performed remarkably well and is generally believed to have prevented severe congestion as the Internet scaled up by six orders of magnitude in size, speed, load, and connectivity. It is also...

Citation Context

...bability, in period . 2 Here, is the peak window size that gives the “average” window of . Hence, a flow-level model of AIMD is Setting in (1) yields the well-known formula for TCP Reno discovered in =-=[37]-=-, [48], which relates loss probability to window size in equilibrium: 1 It should be @I 0 � @�AA packets, where � @�A is the end-to-end loss probability. This is roughly 1 when � @�A is small. 2 This ...

Citation Context

...ol To appear in IEEE/ACM Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], [3], [72], [12]. See [5], [68], =-=[28]-=-, [27], [56], [75], [36], [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability ...

Citation Context

...of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–[35], [39], [41], [42], [44], [46], [50]–[53], [56], [63], [64], [73], [76] (see also, e.g., [30], [40], [43], =-=[57]-=- for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, clarify ideas, and suggest directions, leading to...

Citation Context

...ut also what it does to network queues and how that affects other applications sharing the same queue. We compare the performance of FAST TCP with Reno, HSTCP (HighSpeed TCP [15]), STCP (Scalable TCP =-=[32]-=-), and BIC TCP [75], using their default parameters. 1sIn Section V, we summarize open issues and provide references for proposed solutions. II. MOTIVATIONS A congestion control algorithm can be desig...

Citation Context

...indow control algorithm must be stable, in addition to being fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [31], [42], [35], [50], =-=[46]-=-, [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation of the flow-level...

Citation Context

...el, but it induces certain flow-level properties such as throughput, fairness, and stability. These properties can be understood with a flow-level model of the AIMD algorithm, e.g., [29], [19], [39], =-=[41]-=-. The window wi(t) of source i increases by 1 packet per RTT, 1 and decreases per unit time by xi(t)qi(t) · 1 4 · 2 3 wi(t) packets where xi(t) := wi(t)/Ti(t) pkts/sec. Ti(t) is the round-trip time, a...

Citation Context

...fers. The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 [20] and has gone through several enhancements since, e.g., [21], [58], [47], =-=[18]-=-, [1], [14]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Internet scaled up by six orders of magnitude in size, speed, load, and connectivity....

Citation Context

...lved in data transfers. The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 [20] and has gone through several enhancements since, e.g., =-=[21]-=-, [58], [47], [18], [1], [14]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Internet scaled up by six orders of magnitude in size, speed, load,...

Citation Context

... at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–[35], [39], [41], [42], [44], [46], [50]–[53], [56], [63], [64], [73], =-=[76]-=- (see also, e.g., [30], [40], [43], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, clarify i...

Citation Context

...The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 [20] and has gone through several enhancements since, e.g., [21], [58], [47], [18], =-=[1]-=-, [14]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Internet scaled up by six orders of magnitude in size, speed, load, and connectivity. It i...

Citation Context

... large-scale networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., =-=[43]-=-, [40], [30], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, clarify ideas, and suggest dire...

Citation Context

...a packet-level model, but it induces certain flow-level properties such as throughput, fairness, and stability. These properties can be understood with a flow-level model of the AIMD algorithm, e.g., =-=[29]-=-, [19], [39], [41]. The window wi(t) of source i increases by 1 packet per RTT, 1 and decreases per unit time by xi(t)qi(t) · 1 4 · 2 3 wi(t) packets where xi(t) := wi(t)/Ti(t) pkts/sec. Ti(t) is the ...

Citation Context

...on. Delay-based congestion control To appear in IEEE/ACM Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], =-=[3]-=-, [72], [12]. See [5], [68], [28], [27], [56], [75], [36], [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately...

Citation Context

... Reno, the window control algorithm must be stable, in addition to being fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [31], [42], =-=[35]-=-, [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation of th...

Citation Context

...based solution. Delay-based congestion control To appear in IEEE/ACM Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in =-=[23]-=-, [69], [3], [72], [12]. See [5], [68], [28], [27], [56], [75], [36], [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more...

Citation Context

... The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 [20] and has gone through several enhancements since, e.g., [1], [14], [18], [21], =-=[47]-=-, [58]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Internet scaled up by six orders of magnitude in size, speed, load, and connectivity. It i...

Citation Context

...hese difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], [12], [23], [69], [72]. See [4], [5], [27], [28], [36], [56], [68], =-=[75]-=- for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability both because packet losses in...

Citation Context

...et-level model, but it induces certain flow-level properties such as throughput, fairness, and stability. These properties can be understood with a flow-level model of the AIMD algorithm, e.g., [29], =-=[19]-=-, [39], [41]. The window wi(t) of source i increases by 1 packet per RTT, 1 and decreases per unit time by xi(t)qi(t) · 1 4 · 2 3 wi(t) packets where xi(t) := wi(t)/Ti(t) pkts/sec. Ti(t) is the round-...

Citation Context

... both analytically and experimentally. In Section III-B, we present a mathematical model of the window control algorithm. We prove that FAST TCP has the same equilibrium properties as TCP Vegas [50], =-=[44]-=-. In particular, it does not penalize flows with large propagation delays and it achieves weighted proportional fairness [31]. For the special case of single bottleneck link with heterogeneous flows, ...

Citation Context

...ry differential equation model of TCP/AQM, the dynamics of queueing delay has the right scaling with respect to network capacity. This helps maintain stability as a network scales up in capacity [8], =-=[51]-=-, [53]. In Section III, we lay out an architecture and present an overview of some of the algorithms implemented in our first prototype. Even though the discussion is in the context of FAST TCP, the a...

Citation Context

...air and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–[35], [39], [41], [42], [44], [46], [50]–[53], [56], =-=[63]-=-, [64], [73], [76] (see also, e.g., [30], [40], [43], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand...

Citation Context

...sfers. The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 [20] and has gone through several enhancements since, e.g., [1], [14], [18], =-=[21]-=-, [47], [58]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Internet scaled up by six orders of magnitude in size, speed, load, and connectivity...

Citation Context

... networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], =-=[30]-=-, [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, clarify ideas, and suggest directions, lead...

Citation Context

... Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], [3], [72], [12]. See [5], [68], [28], [27], [56], [75], =-=[36]-=-, [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability both because packet loss...

Citation Context

...stion control To appear in IEEE/ACM Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], [3], [72], [12]. See =-=[5]-=-, [68], [28], [27], [56], [75], [36], [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss ...

Citation Context

...cient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–[35], [39], [41], [42], [44], [46], [50]–[53], [56], [63], [64], =-=[73]-=-, [76] (see also, e.g., [30], [40], [43], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, cla...

Citation Context

...n addition to being fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], [6], [51], [63], =-=[34]-=-, [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important ro...

Citation Context

...el model, but it induces certain flow-level properties such as throughput, fairness, and stability. These properties can be understood with a flow-level model of the AIMD algorithm, e.g., [29], [19], =-=[39]-=-, [41]. The window wi(t) of source i increases by 1 packet per RTT, 1 and decreases per unit time by xi(t)qi(t) · 1 4 · 2 3 wi(t) packets where xi(t) := wi(t)/Ti(t) pkts/sec. Ti(t) is the round-trip t...

Citation Context

... of throughput achieved, but also what it does to network queues and how that affects other applications sharing the same queue. We compare the performance of FAST TCP with Reno, HSTCP (HighSpeed TCP =-=[15]-=-), STCP (Scalable TCP [32]), and BIC TCP [75], using their default parameters. 1sIn Section V, we summarize open issues and provide references for proposed solutions. II. MOTIVATIONS A congestion cont...

Citation Context

...y and a stable design of the flow dynamics. We explain these difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], [12], [23], =-=[69]-=-, [72]. See [4], [5], [27], [28], [36], [56], [68], [75] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estim...

Citation Context

...-scale networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], =-=[40]-=-, [30], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, clarify ideas, and suggest directions...

Citation Context

...mechanisms, see [71], [24]. An initial prototype that included some of these features was demonstrated in November 2002 at the SuperComputing Conference, and the experimental results were reported in =-=[26]-=-. In the following, we explain in detail the design of the window control component. B. Window control algorithm FAST reacts to both queueing delay and packet loss. Under normal network conditions, FA...

Citation Context

...y differential equation model of TCP/AQM, the dynamics of queueing delay has the right scaling with respect to network capacity. This helps maintain stability as a network scales up in capacity [51], =-=[8]-=-, [53]. In Section III, we lay out an architecture to implement our design, and present an overview of some of the algorithms implemented in our current prototype. Even though the discussion is in the...

Citation Context

...actions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], [3], [72], [12]. See [5], [68], [28], [27], [56], [75], [36], =-=[4]-=- for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability both because packet losses in...

Citation Context

...ferential equation model of TCP/AQM, the dynamics of queueing delay has the right scaling with respect to network capacity. This helps maintain stability as a network scales up in capacity [8], [51], =-=[53]-=-. In Section III, we lay out an architecture and present an overview of some of the algorithms implemented in our first prototype. Even though the discussion is in the context of FAST TCP, the archite...

Citation Context

...congestion signals, Fig. 8. Topology with multiple bottlenecks. Fig. 9. FAST with multiple bottlenecks. such as FAST (which uses delay) and Reno (which uses loss), can be very intricate. See [60] and =-=[61]-=- for details and more comments in Section V. 3) FAST in a Network With Multiple Bottlenecks: We simulated a network with two LANs connected by a WAN. The topology is shown in Fig. 8. Three pairs of fl...

Citation Context

...l , then the system converges in finite time. The stability result reported here is limited to local asymptotic stability at a single link with heterogeneous sources and feedback delay is ignored. In =-=[65]-=-, the local stability result is extended to a multilink network in the absence of feedback delay. With feedback delay, local stability can be maintained for the case of a single link, provided the het...

Citation Context

...appear in IEEE/ACM Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], [3], [72], [12]. See [5], [68], [28], =-=[27]-=-, [56], [75], [36], [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability both b...

Citation Context

... of this estimation error is equivalent to modifying the underlying utility functions to favor new flows over existing flows; see [44] and [59]. Methods to eliminate this error are suggested in [44], =-=[49]-=-, and [59]. Propagation delay measurement can be affected by route change from a shorter to a longer path during the lifetime of a connection. Though route change at the timescale of TCP connections i...

Citation Context

...is justified because of self-clocking: within one round-trip time after a congestion window is increased, packet transmission will be clocked at the same rate as the throughput the flow receives. See =-=[66]-=- for detailed justification and validation experiments. A consequence of this assumption is that the link queueing delay vector, , is determined implicitly by the instantaneous window size in a static...

Citation Context

...We explain these difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], [12], [23], [69], [72]. See [4], [5], [27], [28], [36], =-=[56]-=-, [68], [75] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability both because pack...

Citation Context

...a stable design of the flow dynamics. We explain these difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], [12], [23], [69], =-=[72]-=-. See [4], [5], [27], [28], [36], [56], [68], [75] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated t...

Citation Context

...must be stable, in addition to being fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], =-=[6]-=-, [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation of the flow-level design. The theory pla...

Citation Context

...ifferent congestion signals, Fig. 8. Topology with multiple bottlenecks. Fig. 9. FAST with multiple bottlenecks. such as FAST (which uses delay) and Reno (which uses loss), can be very intricate. See =-=[60]-=- and [61] for details and more comments in Section V. 3) FAST in a Network With Multiple Bottlenecks: We simulated a network with two LANs connected by a WAN. The topology is shown in Fig. 8. Three pa...

Citation Context

...lain these difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], [12], [23], [69], [72]. See [4], [5], [27], [28], [36], [56], =-=[68]-=-, [75] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated than loss probability both because packet los...

Citation Context

...evel. The emerging theory of large-scale networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], [33], [64], [53], [8], [52], [73], =-=[11]-=-, [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, cla...

Citation Context

...rage” window of wi(t). Hence, a flow-level model of AIMD is: ˙wi(t) = 1 2 − Ti(t) 3 xi(t)qi(t)wi(t) (1) Setting ˙wi(t) = 0 in (1) yields the well-known 1/ √ q formula for TCP Reno discovered in [48], =-=[37]-=-, which relates loss probability to window size in equilibrium: q ∗ i = 3 2w ∗2 i In summary, (1) and (2) describe the flow-level dynamics and equilibrium, respectively, for TCP Reno. Even though Reno...

Citation Context

...t is extended in [66] to a multilink network; furthermore, global stability at a single link is established in the absence of delay using a Lyapunov argument. These results are summarized in [67]. In =-=[9]-=-, a condition is derived under which a single-source single-link network is globally aysmptotically stable under FAST. i αi (7) (8) IV. PERFORMANCE We have conducted experiments on our dummynet [55] t...

Citation Context

...g fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [31], [42], [35], [50], [46], [76], [44], [41], [6], [51], [63], [34], [19], [39], =-=[33]-=-, [64], [53], [8], [52], [73], [11], [56] (see also, e.g., [43], [40], [30], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a ...

Citation Context

... algorithm (4). In a clean network, the queueing delay maintained by existing FAST flows may be mistaken as part of the propagation delay by new flows that join later, as shown in ns-2 simulations in =-=[59]-=-. The effect of this estimation error is equivalent to modifying the underlying utility functions to favor new flows over existing flows; see [44] and [59]. Methods to eliminate this error are suggest...

Citation Context

...ave different clock resolutions, the sender can estimate the receiver clock period by observing the number of ticks of receiver’s clock during a fixed time interval. A different method is proposed in =-=[16]-=- that does not directly measure the queueing delay. Instead, they measure that actual throughput in the forward direction, and use this measurement in place of W d+q . C. α tuning The parameter α in w...

Citation Context

...s probability and a stable design of the flow dynamics. We explain these difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], =-=[12]-=-, [23], [69], [72]. See [4], [5], [27], [28], [36], [56], [68], [75] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accu...

Citation Context

...t only in terms of throughput achieved, but also what it does to network queues and how that affects other applications sharing the same queue. We compare the performance of FAST TCP with Reno, HSTCP =-=[15]-=-, STCP [32], and BIC TCP [75], using their default parameters. In Section V, we summarize open issues and provide references for proposed solutions. II. MOTIVATIONS A congestion control algorithm can ...

Citation Context

...ased congestion control To appear in IEEE/ACM Transactions on Networking, 2007. An abridged version appears in [25] and an expanded version in [24]. has been proposed, e.g., in [23], [69], [3], [72], =-=[12]-=-. See [5], [68], [28], [27], [56], [75], [36], [4] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately estimated t...

Citation Context

...77]. The experiment and simulation results reported aim to zoom in on specific properties of FAST. These scenarios may be incomplete or unrealistic. Experiments in production networks can be found in =-=[17]-=-, [26]. Other results not presented in this paper due to space limitation are collected in [70] and [79]. A. Testbed and kernel instrumentation Our testbed consists of a sender and a receiver both run...

Citation Context

...ific properties of FAST. These scenarios may be incomplete or unrealistic. Experiments in production networks can be found in [17] and [26]. Other results not presented in this paper are collected in =-=[70]-=- and [79]. A. Testbed and Kernel Instrumentation Our testbed consists of a sender and a receiver both running Linux, that are connected through an emulated router running dummynet under FreeBSD. Each ...

Citation Context

...trol Protocol (TCP) in source and destination computers involved in data transfers. The congestion control algorithm in the current TCP, which we refer to as Reno in this paper, was developed in 1988 =-=[20]-=- and has gone through several enhancements since, e.g., [1], [14], [18], [21], [47], [58]. It has performed remarkably well and is generally believed to have prevented severe congestion as the Interne...

Citation Context

...trol algorithm must be stable, in addition to being fair and efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–=-=[35]-=-, [39], [41], [42], [44], [46], [50]–[53], [56], [63], [64], [73], [76] (see also, e.g., [30], [40], [43], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an ...

Citation Context

...tamp option that is widely implemented in today’s TCP stacks. The calculation is correct even when the sender and receiver’s clocks have a constant offset. If the clock drift is significant, [54] and =-=[62]-=- provide techniques to accurately synchronize the clocks without GPS (global positioning system). If the sender and receiver have different clock resolutions, the sender can estimate the receiver cloc...

Citation Context

...d efficient, at the flow level. The emerging theory of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–[35], [39], [41], [42], [44], [46], [50]–[53], [56], [63], =-=[64]-=-, [73], [76] (see also, e.g., [30], [40], [43], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issue...

Citation Context

...ts with singlebottleneck networks and identical protocol. We conducted ns-2 simulation to study the performance of FAST in more complex environments. The FAST implementation in ns-2 is from CUBIN Lab =-=[78]-=-. To eliminate possible simulation artifacts, such as phase effect, we introduced two-way noise traffic in the simulation, where a certain number of Pareto on–off flows with shape parameter 1.5 were i...

Citation Context

...three kernels are nearly identical. Linux TCP implementation includes all of the latest RFCs such as New Reno, SACK, D-SACK, and TCP high performance extensions. There are two versions of HSTCP [38], =-=[10]-=-. We present the results of the implementation in [38], but our tests show that the implementation in [10] has comparable performance. In all of our experiments, the bottleneck capacity is 800 Mbps—ro...

Citation Context

...f the three kernels are nearly identical. Linux TCP implementation includes all of the latest RFCs such as New Reno, SACK, D-SACK, and TCP high performance extensions. There are two versions of HSTCP =-=[38]-=-, [10]. We present the results of the implementation in [38], but our tests show that the implementation in [10] has comparable performance. In all of our experiments, the bottleneck capacity is 800 M...

Citation Context

...ability and a stable design of the flow dynamics. We explain these difficulties in Section II and motivate a delaybased solution. Delay-based congestion control has been proposed, e.g., in [3], [12], =-=[23]-=-, [69], [72]. See [4], [5], [27], [28], [36], [56], [68], [75] for other recent proposals. Using queueing delay as a congestion measure has two advantages. First, queueing delay can be more accurately...

Citation Context

...e emerging theory of large-scale networks under end-to-end control, e.g., [6], [8], [11], [19], [31], [33]–[35], [39], [41], [42], [44], [46], [50]–[53], [56], [63], [64], [73], [76] (see also, e.g., =-=[30]-=-, [40], [43], [57] for recent surveys), forms the foundation of the flow-level design. The theory plays an important role by providing a framework to understand issues, clarify ideas, and suggest dire...

Citation Context

...three kernels are nearly identical. Linux TCP implementation includes all of the latest RFCs such as New Reno, SACK, D-SACK, and TCP high performance extensions. There are two versions of HSTCP [10], =-=[38]-=-. We present the results of the implementation in [38], but our tests show that the implementation in [10] has comparable performance. In all of our experiments, the bottleneck capacity is 800 Mb/s—ro...

Citation Context

...TCP timestamp option that is widely implemented in today’s TCP stacks. The calculation is correct even when the sender and receiver’s clocks have a constant offset. If the clock drift is significant, =-=[54]-=- and [62] provide techniques to accurately synchronize the clocks without GPS (global positioning system). If the sender and receiver have different clock resolutions, the sender can estimate the rece...

Citation Context

...nously. The data control component determines which packets to transmit, window control determines how many packets to transmit, and burstiness control determines when to transmit these packets [24], =-=[71]-=-. These decisions are made based on information provided by the estimation component. An initial prototype that included some of these features was demonstrated in November 2002 at the SuperComputing ...

Citation Context

...erties of FAST. These scenarios may be incomplete or unrealistic. Experiments in production networks can be found in [17] and [26]. Other results not presented in this paper are collected in [70] and =-=[79]-=-. A. Testbed and Kernel Instrumentation Our testbed consists of a sender and a receiver both running Linux, that are connected through an emulated router running dummynet under FreeBSD. Each testbed m...