N. Cardwell, S. Savage, and T. Anderson, "Modeling the Performance of Short TCP Connections," Tech. Rep., Department of Computer Science and Engineering, Univ. of Washington, 1998, http://www.cs.washington.edu/homes/cardwell/quals.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....limit the data set of interest. In the case of the GridFTP monitoring data, initial results showed that file transfer rates had a strong correlation to file size. Studies of Internet traffic have also revealed that small files achieve low bandwidths while larger files tend to have high bandwidths [4, 16]. This is thought to be primarily due to the startup overhead associated with the TCP slow start mechanism that probes the bandwidth at connection startup. Recent work has attempted to perform class based isolation of TCP flows [35] or startup optimizations [37, 38] to attempt to mitigate this ....

N. Cardwell, S. Savage, and T. Anderson, Modeling the performance of short TCP connections, Technical Report, Computer Science Department, Washington University, Nov. 1998.

....recovery mechanisms. Thus, due to the TCP daunting complexity we note that the current studies focus on specific aspects of TCP performance in restricted simulation scenarios [2, 6, 7, 3, 4, 8, 9] network links on experimental network configurations [5, 10] and selected client to server links [11]. Thus, the analysis presented in the abovementioned studies are constrained by the fact that a real Internet connection is penalized by interfering traffic, which is not captured by simulation models. On the other hand, users navigate the Internet in a random manner. Thus, the quality of service ....

....of the server application that remains blocked in the accept system call. Thus, the impact of server processing time is kept at a minimum. Secondly, the loss of either the SYN segment from client to server or the ACK from server to client produces a retransmission with a three seconds timer [11]. This timer is deterministic because there is no way for the TCP protocol agent neither at the client nor at the server side to achieve an RTT estimate at the connection setup phase. The detection of a retransmission event is thus straightforward, so that anomalous RTT estimation can be filtered ....

S. Savage N. Cardwell and T. Anderson. Modeling the performance of short TCP connections. Available in http://www.cs.washington.edu/homes/cardwell/ quals/quals-paper.ps, October 1998.

....to limit the data set of interest. With the GridFTP monitoring data, initial results showed that file transfer rates had a strong correlation with file size. Studies of Internet traffic have also revealed that small files achieve low bandwidths whereas larger files tend to have high bandwidths [4, 5, 19]. This difference is thought to be primarily due to the startup overhead associated with the TCP start mechanism that probes the bandwidth at connection startup. Recent work has focused on class based isolation of TCP flows [43] and on startup optimizations [45, 46] to mitigate this problem. As a ....

N. Cardwell, S. Savage, and T. Anderson, Modeling the Performance of Short TCP Connections, Technical report, Computer Science Department, Washington University, November 1998.

....be able to differentiate between congestion related, transmission (or random) losses and motion related losses. Short Flows Web browsing and e mail account for a large majority of today s Internet traffic [8, 9] These services usually include the transmission of rather small amounts of data [37]. This means that when the application layer protocol opens a TCP connection for the transfer, there is a very large probability that the whole transfer is completed while the TCP sender is still in the slow start phase. Therefore, the TCP connection never manages to fully utilize the available ....

N. Cardwell, S. Savage, and T. Anderson, "Modeling the Performance of Short TCP Connections," Technical Report, Computer Science Department, Washington University, Nov. 1998.

....RTT s it increases its rate by r s , and responds to loss detection by decreasing its rate to be j 2 (0; 1) times the previous rate. This behaviour is a simplification of the congestion response of a TCP flow. We need to use this method of global modelling, rather than using known TCP equations [6, 7, 3, 4, 1]. Those relate loss rate to throughput for an indiviual source, with no concept of the number of bottlenecks a flow experiences. Suppose ff = ff l for all l = 1 : L, namely the drop bias is the same on every link. Then it can be shown that the distribution of long term rates x s can be ....

....network all flows have a round trip time f , where f 2 [ b ; g ] Let x b and x g be the throughput received by a blue and green flow respectively in the ABE environment. In the flat best effort scenario they would have throughput x f . Assuming the established loss throughput formula [6, 7, 3, 4, 1] holds, x b = C b p q b ; x g = C g p q g and x b = C f p q f where q b , q g , q g are the blue, green, and flat loss ratios respectively, and C is a constant. Assume we can characterise the dropping at the router by a nondecreasing convex function OE(x) where OE(0) 0 and ....

N. Cardwell, S. Savage, T. Anderson. Modeling the Performance of Short TCP Connections. University of Washington. Oct, 1998.

....to come up with a general TCP model taking into account of all the details. On the other hand, if we know the distinguishing feature of a TCP connection, we can explore it to improve the accuracy of approximation. For example, if the connection is short, we need to consider the slow start phase [2]. In summary, the model is successful in that its analytical results are usually close to simulation results, and its development also gives us insights about on the performance of a single TCP connection under cross traffic. 4 Aggregate Behavior of TCP Connections Model Simplification Section ....

N. Cardwell, S. Savage, and T. Anderson. Modeling the Performance of Short TCP Connections. Techical Report.

....the server. The number of trials in each experiment is so determined to take both the time that a single trial takes and the stability of the results into account. Since the retrieval latency exhibits a large variance due to the various ways that the packet loss can affect the progress of a flow [9], a high stability (i.e. error 5 ) of the results can only be achieved with at least 10,000 trials, which is infeasible in terms of the simulation time 4 . Fortunately, the number of trials on the magnitude of hundreds is good enough to capture the trend of the retrieval latency. The number ....

....with the network offered load equal to 90 may take around 200 seconds to complete. 5 In generic HTTP 1. 0 [5] the client opens up a new TCP connection to retrieve each object embedded in a Web page, which results in the median and average transfer sizes to be 2 3KB and 8 12KB, respectively [9]. 6 In HTTP 1.1, a persistent TCP connection can be used to retrieve several objects [12] Pipelining allows the client to pipeline all the requests for the embedded objects in a page and the server to pipeline all the corresponding responses. The resulting average transfer size of a TCP ....

[Article contains additional citation context not shown here]

N. Cardwell, S. Savage, and T. Anderson, Modeling the performance of short TCP connections, Technical Report, Computer Science Department, Washington University, Nov. 1998.

....[2] are better suited for FIFO drop tail queueing [2] Also, for short transfers, the cwnd values are expected to be small and the flows usually do not use a significant portion of the path s bandwidth. A binomial model is thus well suited for modeling the total number of losses suffered by a flow [1]. Finally, we assume that the time to transmit all the packets in a round is smaller than the duration of the round and that the duration of a round is independent of the window size. The amount of data transferred is assumed to be arbitrary allowing on one hand for extremely short transfers which ....

N. Cardwell, S. Savage and T. Anderson, "Modeling the performance of short TCP Connections," University of Washington, Seattle, Oct 1998.

....flows. Our focus here has been on using isolation to improve predictability of the service and fairness of TCP flows of different size. Most TCP studies consider very long (or infinite) TCP connections, and focus on characterizing the steady state transfer throughput. Only few recent TCP studies [58], 29] 59] have started to investigate short flows, which comprise most of the current Internet flows [15] However, to our knowledge, the interaction among short and long TCP flows has not been studied. VI. CONCLUSIONS AND FUTURE WORK Using control theoretic arguments and extensive simulations, ....

N. Cardwell, S. Savage, and T. Anderson, "Modeling the Performance of Short TCP Connections," Tech. Rep., Department of Computer Science and Engineering, Univ. of Washington, 1998, http://www.cs.washington.edu/homes/cardwell/quals.

....[2] are better suited for FIFO drop tail queueing [2] Also, for short transfers, the cwnd values are expected to be small and the flows usually do not use a significant portion of the path s bandwidth. A binomial model is thus well suited for modeling the total number of losses suffered by a flow [1]. Finally, we assume that the time to transmit all the packets in a round is smaller than the duration of the round and that the duration of a round is independent of the window size. The amount of data transferred is assumed to be arbitrary allowing on one hand for extremely short transfers which ....

N. Cardwell, S. Savage and T. Anderson, "Modeling the performance of short TCP Connections," University of Washington, Seattle, Oct 1998.

....average transfer delay. The graph is the result of NS 1.4 [4] simulations at different uniform loss rates, using 20 packet transfers, TCP Tahoe [3] no delayed acks, and a 0.1 second round trip propagation time. These results are similar to those produced by Cardwell s model for short connections [5]. There is nothing very surprising in Figure 1: as the loss rate increases, TCP slows down. Hidden by the averages, however, are significantly skewed distributions, illustrated by Figure 2. Each of the plots shows the cumulative distribution of the time required to complete simulated 20 packet ....

.... The following function approximates the average window size (w) that TCP uses when faced with a particular average loss rate (l) w = 0:87 p l (1) This formula is adapted from Floyd [6] and Mathis et al. 7] more detailed approximations can be found in Padhye et al. 8] and Cardwell et al. [5]. Equation 1 can be viewed in two ways. First, if the network discards packets at a rate independent of the sender s actions, Equation 1 describes how the sender will react. Second, if the network can store only a limited number of packets, Equation 1 indicates the loss rate the network must ....

Neal Cardwell, Stefan Savage, and Tom Anderson, "Modeling the Performance of Short TCP Connections," Tech. Rep., University of Washington, 1998.

....TCP connection for Web retrieval. Both HTTP 1.0 and HTTP 1.1 are studied in our experiments. In generic HTTP 1. 0, the client opens up a new TCP connection to retrieve each object embedded in a Web page [6] which results in the median and average transfer sizes to be 2 3KB and 8 12KB, respectively [9]. Correspondingly, our experiments on HTTP 1.0 are conducted on file sizes 3KB and 10KB. The key differences between HTTP 1.0 and HTTP 1.1 are persistent connection and request pipelining. In HTTP 1.1, a persistent TCP connection can be used to retrieve several objects [12] Pipelining allows the ....

....Number of trials The number of trials in each experiment is so determined to take both the time that a single trial takes and the stability of the results into account. Since the retrieval latency exhibits a large variance due to the varied ways that packet loss can affect the progress of a flow [9], a high stability (i.e. error 5 ) of the results can only be achieved with at least 10,000 trials, which is infeasible in terms of simulation time (a typical trial for 30KB file with network offered load 90 may take around 200 seconds to complete) Fortunately, the number of trials on the ....

[Article contains additional citation context not shown here]

N. Cardwell, S. Savage, and T. Anderson, Modeling the performance of short TCP connections, Technical Report, Computer Science Department, Washington University, Nov. 1998.

No context found.

N. Cardwell, S. Savage, and T. Anderson, "Modeling the Performance of Short TCP Connections," Tech. Rep., Department of Computer Science and Engineering, Univ. of Washington, 1998, http://www.cs.washington.edu/homes/cardwell/quals.

No context found.

N. Cardwell, "Modeling the performance of short TCP connections," unpublished.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC