M. Allman. TCP Byte Counting Refinements. Computer Communication Review, 29(3), July 1999.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Gb sec links; and yet there two protocols, UDP and TCP, that govern most of this communication. For this reason, many have proposed modifications and specializations to UDP or TCP to better serve the needs of applications. A literature search for proposed modifications quickly returns many results [3, 4, 5, 6, 10, 11, 13, 16, 18, 20, 23, 25, 28, 30, 31]. This work was funded by generous grants from NSF (DGE 9616736) and DARPA (F30602 00 2 0565) Correspondence concerning this paper may be sent to ely cs.washington.edu. Because the networking stack is traditionally part of the operating system, most of these modifications were developed and ....

M. Allman. TCP byte counting refinements. Computer Communications Review, 29(3), July 1999.

....set of possible analyses as selectable options, users are free to modify this program or to write their own analyzer to study the trace data in a manner most relevant to their interests. 3 NETWORK ANALYSIS Much has been written about various ways to evaluate networking protocols, especially TCP [3][1] FKT has been used to investigate the TCP, UDP, IP, and LAMP[4] protocol stacks in Linux. Probes were inserted at the entry and exit points for all the major kernel functions in these stacks, and then two user level test programs were run: a ping pong program to test round trip latency, and a ....

M. Allman and A. Falk. On the eective evaluation of tcp. Computer Communication Review, 29(5):59-70, October 1999.

....information is sent separately from data via a separate signaling protocol; out of order packets are retransmitted rather than being retained by the receiving side, etc. 5 PERFORMANCE Much has been written about various ways to evaluate networking protocols in general and TCP in particular [5] [4] To date we have performed very limited testing of LAMP vs. TCP by using two benchmarks, one to measure throughput, the other to measure latency or response time. These tests were all run on 600 MHz Pentium III processors having 256 Mbytes of memory that are interfaced via DEC DS21140 tulip ....

M. Allman and A. Falk. On the eective evaluation of tcp. Computer Communication Review, 29(5):59-70, October 1999.

....segment. There are two obvious solutions: either modify the congestion control mechanisms to operate at byte granularity or guarantee that segment level granularity is always respected. The first solution is virtually identical to the byte counting modifications to TCP discussed in [All98, All99] If cwnd is not incremented by a full SMSS, but only proportional to the amount of data acknowledged, then ACK division attacks will have no effect. The second, perhaps simpler, solution is to only increment cwnd by one SMSS when a valid ACK arrives that covers the entire data segment sent. As ....

Mark Allman. TCP byte counting refinements. Computer Communications Review, 29(3), July 1999.

No context found.

M. Allman. TCP Byte Counting Refinements. Computer Communication Review, 29(3), July 1999.

No context found.

M. Allman. On the Generation and Use of TCP Acknowledgements. Computer Communication Review, 28(5), October 1998.

No context found.

M. Allman. TCP Byte Counting Refinements. Computer Communication Review, 29(3), July 1999.

No context found.

M. Allman. On the Generation and Use of TCP Acknowledgements. Computer Communication Review, 28(5), October 1998.

No context found.

M. Allman. TCP Byte Counting Refinements. Computer Communication Review, 29(3), July 1999.

No context found.

M. Allman. On the Generation and Use of TCP Acknowledgements. Computer Communication Review, 28(5), October 1998.

No context found.

M. Allman. TCP Byte Counting Refinements. Computer Communication Review, 29(3), July 1999.

No context found.

M. Allman. On the Generation and Use of TCP Acknowledgements. Computer Communication Review, 28(5), October 1998.

....measurement is completed. As outlined in [PMAM98] a mesh of hosts, such as NIMI, has good scaling properties. Utilizing a mesh of n hosts we can measure O(n 2 ) network paths. Our investigation spanned nearly 900 network paths yielding results without strong biases based on the network. See [AF99] for a comparison of measurement techniques. We installed two scripts on the hosts in the NIMI infrastructure to conduct our tests. Each script performs two data transfers, as well as recording various other information about the test. For instance, we record the time each measurement is ....

Mark Allman and Aaron Falk. On the Effective Evaluation of TCP. Computer Communication Review, 29(5):59--70, October 1999.

.... (e.g. using a larger initial congestion window, as outlined in section 8) Also, our approach allows for the measurement of properties of a large number of real clients (e.g. which TCP options are supported in their stacks) While our approach is not without flaw, we believe (as outlined in [AF99] there is no perfect way to assess Internet behavior and therefore believe that this survey can provide valuable insight into the performance of the web as seen from the user s perspective (since all traffic studied comes from web browsing as it happens in the wild ) Throughout this study we ....

....SACK is lagging behind the percentage of hosts supporting the option. This indicates that a number of the web crawlers that hit our server many times per month do not support the SACK option yet. We believe the SACK deployment shown in this plot is consistent with the recommendation made in [AF99] that SACK should be a part of all TCP investigations, as SACK is clearly steadily being deployed in the Internet. 5 Round Trip Times This section focuses on examining the distribution of round trip times (RTT) between the server and the clients. We used tcptrace to produce the average and ....

[Article contains additional citation context not shown here]

Mark Allman and Aaron Falk. On the Effective Evaluation of TCP. Computer Communication Review, 29(5):59--70, October 1999.

.... (e.g. using a larger initial congestion window, as outlined in section 9) Also, our approach allows for the measurement of properties of a large number of real clients (e.g. which TCP options are supported in their stacks) While our approach is not without flaw, we believe (as outlined in [AF99] there is no perfect way to assess Internet behavior and therefore believe that this survey can provide valuable insight into the performance of the web as seen from the user s perspective (since all traffic studied comes from web browsing as it happens in the wild ) Throughout this study we ....

....SACK is lagging behind the percentage of hosts supporting the option. This indicates that a number of the web crawlers that hit our server many times per month do not support the SACK option yet. We believe the SACK deployment shown in this plot is consistent with the recommendation made in [AF99] that SACK should be a part of all TCP investigations, as SACK is clearly steadily being deployed in the Internet. 6 Round Trip Times This section focuses on examining the distribution of round trip times (RTT) between the server and the clients. We used tcptrace to produce the average and ....

[Article contains additional citation context not shown here]

Mark Allman and Aaron Falk. On the Effective Evaluation of TCP. Computer Communication Review, 29(5):59--70, October 1999.

....algorithm that determines when an ACK must be generated. Therefore, to obtain 2 full sized segments worth of data, three segments must be received. Generating fewer ACKs hinders the growth of the congestion window, and therefore the performance TCP and HTTP are able to attain [Pax97, PAD 99, All98] Our fixes corrected this problem so that ACKs are correctly generated for every second full sized data segment that arrives. 2.2 Larger Initial Window In some of the experiments outlined in this paper, we tested an experimental TCP modification that increases the size of the initial ....

Mark Allman. On the Generation and Use of TCP Acknowledgments. Computer Communication Review, 28(5), October 1998.

....using NIMI like environments is that often TCP researchers need to change the TCP implementation in the operating system. For instance, researchers may want to investigate a particular change to TCP s slow start algorithm that requires changing the TCP implementation in the kernel (e.g. All98, All99] This may not be possible on a preexisting mesh of hosts, as the kernel change may interrupt or invalidate experiments being conducted by other researchers. Also, setting up a mesh of hosts all running a custom built kernel can be quite difficult and time consuming. One way to mitigate this ....

....congested competing TCP flows will back off, while flow F will not. Thus, flow F is stealing bandwidth from the competing flows and using a disproportionate amount of the bandwidth. Several methods for measuring TCP s fairness to competing traffic have been used in various studies [Jai91, BP95, All99] ffl Router Queue Length. Some TCP experiments can benefit from measurements of the router queue length (again, easy in a simulator but non trivial in Internet tests) The length of the router queue can have an impact on the end to end delay seen by an application. For instance, interactive ....

Mark Allman. TCP Byte Counting Refinements. Computer Communication Review, 29(3), July 1999.

....of using NIMI like environments is that often TCP researchers need to change the TCP implementation in the operating system. For instance, researchers may want to investigate a particular change to TCP s slow start algorithm that requires changing the TCP implementation in the kernel (e.g. All98, All99] This may not be possible on a preexisting mesh of hosts, as the kernel change may interrupt or invalidate experiments being conducted by other researchers. Also, setting up a mesh of hosts all running a custom built kernel can be quite difficult and time consuming. One way to mitigate ....

.... unchanged TCP (figure 1) denoted TCP0 , and two experimental changes to TCP (figures 2 and 3) denoted TCP1 and TCP2 respectively, in a wide variety of scenarios (various transfer sizes and various router queue lengths) These figures and a discussion of the proposed mechanisms can be found in [All98] These figures are used in this paper as illustration, so the particular changes being proposed are irrelevant. The proposed changes shown are fairly minor and were not expected to radically change TCP s behavior. However, as shown in figure 2, the changes proposed in TCP1 drastically alter the ....

Mark Allman. On the Generation and Use of TCP Acknowledgments. Computer Communication Review, 28(5), October 1998.

....an entire segment. There are two obvious solutions: either modify the congestion control mechanisms to operate at byte granularity or guarantee that segment level granularity is always respected. The first solution is virtually identical to the byte counting modifications to TCP discussed in [All98, All99] If cwnd is not incremented by a full SMSS, but only proportional to the amount of data acknowledged, then ACK division attacks will have no effect. The second, perhaps simpler, solution is to only increment cwnd by one SMSS when a valid ACK arrives that covers the entire data segment ....

Mark Allman. On the generation and use of TCP acknowledgments. Computer Communications Review, 28(5), October 1998.

....OH 44135 mallman grc.nasa.gov Abstract TCP s delayed acknowledgment algorithm has been shown to hurt TCP performance. One method of gaining the performance lost by reducing the number of acknowledgments sent is to use a limited byte counting algorithm. However, we show that as outlined in [All98] limited byte counting is too aggressive in some situations. This paper defines an appropriate byte counting algorithm to fix this aggressiveness. This paper shows that appropriate byte counting is a better overall algorithm. In addition, a scaled version of the appropriate byte counting ....

....for increasing TCP s initial window to 3 or 4 segments (depending on the segment size) Increasing the initial window provides the most benefit for short flows and low bandwidth network paths. Several researchers have studied the impact of using a larger initial window [AHO98, PN98, SP98] All98] suggests using a limited byte counting (LBC) algorithm to mitigate the impact of delayed acknowledgments on TCP performance. TCP s delayed acknowledgment strategy [Bra89] has been shown to reduce the performance of TCP transfers by slowing the growth of cwnd [Pax97, All98, PAD 99] The ....

[Article contains additional citation context not shown here]

Mark Allman. On the Generation and Use of TCP Acknowledgments. Computer Communication Review, 28(5), October 1998.

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