PDF - University of Wisconsin–Madison
Contents
1. likely to govern on un congested paths The two TCP variants that we are concerned with in this lab are TCP Reno and TCP Vegas TCP Reno was one of the first of the Jacobson based ver sions of TCP Reno is considered an agressive ver sion of TCP since it is always probing for additional available capacity until a loss event takes place in which case it backs off multiplicative decrease and recovers the lost packets The New Reno and SACK versions of TCP improved the ability to quickly re cover from losses of multiple packets which is not uncommon Collectively all of these variants of TCP attempt to control congestion in the network An alternative approach is to attempt to avoid congestion all together The idea is that if there is some signal that indicates that congestion is starting then adjust the sending rate to keep it from overload ing the path So the question is how might a sender infer congestion other than by simply packet loss which is what is used by TCP Reno The answer is that when a link begins to be overloaded packet will enqueue in the congested router s output buffers The effect of this enqueuing is to lengthen the RTT of data ACK pairs since packets now have to spend time in the queue before being transmitted The direct implication of lengthened RTT values is a de crease in end to end throughput since fewer packets will be sent over a given period of time The basis for the Vegas version of TCP is t2. CS 638 Lab 6 Transport Control Protocol TCP Joe Chabarek and Paul Barford University of Wisconsin Madison jpchaba pb cs wisc edu The transport layer of the network protocol stack layer 4 sits between applications layers 5 7 and the network layer 3 and below The most basic capability that is enabled by transport is multiplex ing the network between multiple applications that wish to communicate with remote hosts Similar to other layers of the network protocol stack transport protocols encapsulate packets with their own header before passing them down to layer 3 and decapsulate packets before passing them up to applications The most simple transport protocol is the User Datagram Protocol UDP UDP provides a mul tiplexing demultiplexing capability for applications but not much more Most significantly UDP pro vides no guarantees for reliability which is unac ceptable for applications like a file transfer Web or Email in which people would very much like to get complete files For this and other reasons the Trans port Control Protocol TCP is the most widely used transport protocol in the Internet today Over 70 of the traffic on the Internet today is carried by TCP TCP is complicated and there are actually many different versions of the protocol There are literally thousands of academic research papers on TCP and TCP variants many IETF RFC s on TCP and en tire books written about TCP s implementation a
3. hat if a decrease in throughput is measured then send rate should be adjusted to stay just below the rate that led to the decrease This approach to managing congestion is called congestion avoidance since it does not rely on packet loss to adjust send rate 2 1 TCP Reno TCP Reno has all the required elements of a reliable connection oriented full duplex protocol which implements flow control and conjestion con trol For flow control a sliding window algorithm is used to ensure that the sender does not overrun the receiver For conjestion control TCP Reno uses Ja cobson Karls to compute the estimated round trip time and the RTO It also included slow start ad ditive increase multiplicative decrease and the fast retransmit fast recovery mechanisms for managing CWND TCP Reno is the default TCP protocol variant used in the FC6 STD image so you don t have to do anything special to use Reno in your experiments You will however need a mechanism for generating traffic that will then utilize the version of TCP Reno running on the Schooner hosts Note that traffic is generated by applications or in our case traffic gen eration utilities those used in this lab are described below 2 2 TCP Vegas Like Reno TCP Vegas has all the required ele ments of a reliable connection oriented full duplex protocol which implements flow control and conges tion management However instead of reacting to congestion in an after the fact fa
4. hooner to specify a loss rate and or a propagation delays on the link between sender and receiver Report of each experiment in your lab notebook Detailed tasks include 1 Using TCP Reno ttcp and Wireshark transfer each of the three files from PC1 to PC2 config ured with a loss probability of 0 1 on the link and record the sequence number versus time throughput and RTT for the connection 2 Using TCP Reno ttcp and Wireshark transfer each of the three files from PC1 to PC2 config ured with a loss probability of 0 1 and a RTT de lay of 30ms on the link and record the sequence number versus time throughput and RTT for the connection 3 Using TCP Vegas ttcp and Wireshark transfer each of the three files from PC1 to PC2 config ured with a loss probability of 0 1 on the link and record the sequence number versus time throughput and RTT for the connection Using TCP Vegas ttcp and Wireshark transfer each of the three files from PC1 to PC2 config ured with a loss probability of 0 1 and a RTT de lay of 30ms on the link and record the sequence number versus time throughput and RTT for the connection
5. ibed below Load the standard FC6 STD disk image and include the RPM for ttcp along with the list of RPMs you have been using for WiresharkKNOTE A lot of data will be generated in this lab so make sure that you have a method for organizing the data before you actually gather it so you don t mix up traces from different experiments 4 1 Baseline Performance Characteris tics We will begin by examining throughput and RTT s and taking a look at sequence numbers in packets to get an idea of how TCP Reno and TCP Vegas each perform Transfer three different sized files of your own choosing from about 100Bytes to 1MB with ttcp from PC1 to PC2 and generate graphs for the statistics that we listed above Wire shark should enable you to do this Report the re sults of the experiments in your lab notebook De tailed tasks include 1 Using TCP Reno ttcp and Wireshark trans fer each of the three files from PC1 to PC2 and record the sequence number versus time for throughput and RTT for the connection 2 Using TCP Vegas ttcp and Wireshark trans fer each of the three files same sizes as before from PC1 to PC2 and record the sequence num ber versus time throughput and RTT for the connection These measurements and statistics will serve as a baseline for performance in an uncongested environ ment 5 Performance in Congested Condi tions Next run a set of tests using the same file sizes as above this time using Sc
6. ion on a link 3 2 1 ttcp Ttcp is a simple tool that starts a TCP or UDP flow between two systems and reports some basic statis tics about the flow A few other features are also available in the utility see the man page for these additional options To initiate a transfer you need a receiver and a transmitter The receiever is started on one machine with ttcp s r on the transmit ting machine the command is ttcp s t IPAddress NOTE use this utility to generate the streams that you will visualize in Wireshark for your experiments 3 2 2 Harpoon Harpoon is a sophisticated tool which can mimic the type and intensity of traffic on a network At its core Harpoon is a tool that runs on a two systems clients and servers The clients determine how and when to request file transfers The servers send data to the clients based on what has been requested An easy way to think about this is as an automated Web environment The important aspects of Harpoon are that the client request streams can be automatically configured to generate traffic that it identical to what has been observed in an actual network The Harpoon software distribution comes with an extensive user manual that describes installation and configuration procedures A link to this manual is provided in the Pre Lab Harpoon is available on the standard FC6 STD Operating System image that you will use for this lab So you will only need to familiarize yourself with ho
7. nd its behavior TCP includes many features such as bi directional connection management ACK based reli ability receiver flow control and congestion control all of which make it an attractive choice for a wide va riety of applications An easy way to think of what TCP does beyond multiplexing demultiplexing is that it governs how and when packets are trans mitted between a sending host and receiving host As such it has a very big influence on the transmis sion performance as measured by either throughput bytes transmitted over a specified time period or response time elapssed time from transmission start to transmission end The flow of data between hosts is determined by windows which dictate how many packets can be in flight between sender and receiver at any point in time 1 Overview and Objectives Unlike prior labs the focus of lab 6 is is on learning more about experimental tools and on ob serving the behavior of the various mechanisms that are part of TCP The reason for this is because in moving from layer 3 to layer 4 we are moving away from the network per se and into end hosts Further more network administrators usually don t spend a lot of time messing around with TCP since there is no programming or management interface to TCP on end hosts The exception is content providers who may make tweaks in an attempt to get better performance on large file transfers In terms of experimental tools a focus of thi
8. s lab is on learning about traffic generation Since the transport layer is concerned with facilitating the transmission of packets between two end hosts hope fully it is clear that we need tools for transmiting data in order to observe how TCP behaves On one hand we could simply have you generate traffic by using an application like the Web and then tracking packet traffic generated by your browsing However it would be difficult to have you do enough browsing by hand to generate traffic sufficient to cause conges tion on even low bandwidth links which is essential in order to observe key aspects of TCP behavior Another tool that you will use in this lab is the Click Modular Router Click is a utility that enables high performance personal computers to operate like a network device Click offers a simple programming framework that you can implement many different kinds of new functions on top of a high performance packet processing system This is a capability that is extremely useful for lab experiments since com modity routers and switches don t offer programming APIs Wireshark will be used to observe the details of TCP behavior Since TCP governs the transmission of packets by end hosts the focus of the lab config urations will be to generate different kinds of traffic that exercise different aspects of TCP s algorithms In particular you will be able to observe the details of TCP connection setup and tear down flow con
9. shion like Reno i e when loss takes place Vegas attempts to avoid con gestion through careful measurement and manage ment of send rate While the ideas behind Vegas are interesting and certainly make sense it can be the case that Vegas is unsuccessful at avoiding conges tion and packet loss takes place In this case Vegas default behavior in terms of loss recovery and win dow management is identical to TCP Reno TCP Vegas along with a few other major TCP variants are distributed with the linux kernel Mod ern 2 6 kernels implement a modular framework which makes changing from TCP Reno to TCP Ve gas is relatively painless See the commands below on how to find the current TCP algorithm in use and to change beteen Reno and Vegas for your experiments There are a number of differences between Reno and Vegas See PreLab readings that you will be measuring NOTE It may be easier to complete all of the tasks experiments with Reno described below then change to Vegas to conduct the remainder of the tasks experiments sh 3 1 sysctl net ipv4 tcp_congestion_control net ipv4 tcp_congestion_control reno sh 3 1 sysctl w net ipv4 tcp_congestion_control vegas net ipv4 tcp_congestion_control vegas sh 3 1 sysctl net ipv4 tcp_congestion_control net ipv4 tcp_congestion_control vegas sh 3 1 sysctl w net ipv4 tcp_congestion_control reno net ipv4 tcp_congestion_control reno 3 Experimental Tools In this section brief over
10. that enables you to configure a PC to behave as a high performance packet processing system The ba sic primitive in Click is called an element A set of elements can be composed in Click to create a packet processing device Many different kinds of devices can be created including routers switches firewalls intrusion detection systems etc Click is easy to use and has a growing library of preconfigured elements A Click of elements are composed in a configu ration file that is read by the Click engine There will be three different configurations used in the ex periments that enable you to create different condi tions for your experiments The first configuration is a standard router with drop tail queues The sec ond configuration implements a RED queue The third configuration delays each packet for 50us with a drop tail queue In each configuration file you will need to change the strings PC1_IP PC2_IP PC3_IP PC8 IP MAC1 MAC2 MAC3 MAC4 MAC5 MAC6 MAC7 MAC8 to the appropriate values NOTE In the interest of time Click will not be used for the required tasks of this lab However you are welcome and encouraged to install it and use it in your experiments if you are interested 4 Tasks Create Topology 1 described in the Pre Lab and determine where to gather measurements for your tests NOTE In the interest of time topology 2 will NOT be used in this lab This topology will be used in all of the experiments descr
11. trol reliability and congestion control Before you pro ceed to the lab description take a minute and think about how you might configure the lab in order to observe these chacteristics We will experiment with two different versions of TCP in this lab TCP Reno and TCP Vegas Finally TCP is entirely implemented on end hosts This means that you can examine how TCP is implemented e g in open source software make tweaks to existing implementations or even create your own implementation While we will not be do ing any TCP hacking in this lab you may be inter ested in doing this at some point in the future Upon completion of the lab exercises below stu dents will be able to e Understand the basic packet transmission behav ior of TCP e Distinguish the differences between TC P Reno and TCP Vegas e Configure and operate simple traffic generators 2 TCP Variants As mentioned above there are many differ ent variants of the Transmission Control Protocol TCP In most cases the primary differentiating feature of a TCP variant is its congestion control mechanism The reason for this is that the con gestion control mechanism has a large impact on performance defined in terms of either throughput or response time Remember the congestion con trol mechanism governs the size of CWND which on paths with congestion will almost always determine the number of unacknowledged packets in flight as opposed to RWND which is
12. types and think about what they are showing you You can also use a Wireshark display filter to limit the traffic you are graphing You are encouraged to play around with the Wireshark GUI to gain familiarity with this observation technique Be sure to ask questions early if you need help 3 2 Generating Traffic In order to examine the behavior of the TCP Reno and Vegas protocols we must be able to gen erate packet traffic This is normally done by users accessing application such as Email or the Web that result in files being transfered between a client and a server However having real users accessing real ap plications in lab experiments is cumbersome and does not lend itself to repeatability or to experiments that demand large traffic loads e g for experiments on how TCP behaves under congested conditions since hundreds or thousands of users who be required You will be using two tools for generating traf fic in this lab The first is ttcp which is a simple tool that generates a single TCP or UDP stream We will use ttcp with Wireshark to gather statistics about different implementations of TCP However ttcp does not address our need to recreate traffic that is a composite of perhaps thousands of users For this purpose the Harpoon traffic generator will be used Harpoon is a sophisitcated tool which we will use in our experiments to generate a sufficient amount of representative background traffic that to create congest
13. util ity called dummynet which is used to augment traffic flows with different characteristics By specifying a bandwidth dummynet can enable an output link that has 1 Gbps native capacity to behave as if it has e g 1 Mbps The implication is that with dummynet we can create a test environment in which it is much easier to create congested conditions Alternatively we can use dummynet to simply specify a loss rate or RTT on a link In this case we need not even bother with creating an input load that exceeds an output link s capacity in order to get a sending host to react to congestion However this approach is not nearly as realistic and will result in behavior that is not very similar to what would be seen in real networks dummynet is available by default in Schooner through the standard network configuration GUI which by clicking on a link enables you to specify link charac teristics including bandwidth capacity and loss rate When you set any sort of traffic shaping com ponent delay loss possibly bandwitdh Schooner interposes a hidden node on the link running dum mynet At experiment swapin the traffic shaping node will use the parameters in your NS file Dur ing experimentation you can use the Modify Traffic Shaping link on the experiment webpage which will allow you to dynamically change the shaping param eters of your experiment 3 4 The CLICK Modular Router CLICK is a utility and programming framework
14. views of the tools that will be used in this lab are provided More details can be found in the readings specified in the Pre Lab 3 1 Measuring and Observing Packet Traffic In a laboratory test environment one of the ma jor goals while running experiments is to observe traffic without disturbing the performance of the system To this end we describe a standard tech nique for measuring packet traffic Remember how ever that another critical task in conducting exper iments with network traffic is to identify useful van tage points in the network i e where you should place your instrumentation for measuring traffic Wireshark a tool that you have used in prior ex periments is one of a number of applications that uses a system library named libpcap to capture packets The library efficiently creates copies pack ets as they are received from the network card set in promiscuous mode and sends them to user pro grams Perhaps the most well know application that uses libpcap as a mechanism for creating a stream of packet copies is Tepdump which is utility available on most systems Wireshark is a popular tool because of its graph ical user interface and easy to use features For our exercises we will be using a graphing module to ob serve the sequences of packets generated by TCP variants To use the graphing module find the TCP Stream Graph entry in the Statistics menu of the Wireshark GUI Try each of the graph
15. w to configure and run Harpoon in your experiments NOTE In the interest of time Harpoon will not be used for the required tasks of this lab However you are welcome and encour aged to install it and use it in your experi ments if you are interested 3 3 Simulating Adverse Network Condi tions The Initial experiments that you will run will pri marily be concerned with observing TCP in ideal conditions i e with no congestion on the path be tween clients and servers However we are also in terested in how the two different variants of TCP behave when a link is congested and packet loss oc curs Normally to create congestion the aggregate demand on a set of input links must exceed the ca pacity of an output link For example if you are using a router in an experiment with 1 Gbps links then the input on some set of input links where the set must be at least size 2 must be greater than 1 Gbps In real life it takes a lot of clients to generate 1 Gbps of traffic For example the average traffic rates for the entire UW Madison campus is usually only several hundred Mbps The good news is that we have several options in Schooner in terms of how we experiment with congestion Using Schooner we actually have the ability to specify different link characteristics Three charac teristics that we are particularyly interested in are bandwidth propagation delay and packet loss rate There is a popular BSD based operating system
Download Pdf Manuals
Related Search