Exercise #3

Flow Behavior, Isolation and Service Differentiation in TCP

What you will get out of this exercise:

• Exposure to some traffic generation tools, like ttcp and iperf
• Hands on use of some traffic capture tools like tcpdump
• Use of some TCP analysis tools
• Some tweaking of packet handling rules using firewall tools
• A detailed, packet-level understanding of TCP behavior as well as the impact on TCP of the queueing discipline used in routers.

• Sally Floyd's RED page. You don't need to be familiar with all of these resources, but this is a good reference library for practical queue management disciplines.
• The classic paper on RED
• Cisco documentation on (W)RED
• Network throughput testing tools: ttcp or iperf.
• Analysis tool of tcpdump files: tcptrace ( patched xplot that produces correct Postscript files)
• Setting TOS bits in packets
• man pages for tcpdump, iptables and whatever other commands you might need

In addition, you will be asked to submit data files and other kinds of files. To do this, put your report and all required files into a directory named ex3. Then tar and gzip this directory into a file called ex3.tar.gz and submit it electronically. (Read the documentation for the Unix "tar" command on how to do this).

Even if you are not able to complete the exercise, be sure to submit a partial lab report. You will get partial credit for your efforts.

This exercise will involve testing TCP on a representative traffic mix. For this, you will use the following topology:

Setting up the interfaces

You will need to set up an address assignment plan for the topology. Setup the first two bytes of all addresses according to the last four digits of your USC ID. If your USC ID is abcdefghij, then all addresses must begin with 1gh.ij. For example, if your USC ID is xxxxxx1234, then all addresses must be in the form of 112.34.x.x.

Answer these questions in your lab report.

1. How many subnets are there in your topology?
2. As with the routing exercise, list all the subnets and the interface addresses.

Setting up routing

Now, you will need to set up routing on the given topology. To do this, read the zebra documentation to figure out how RIP can be configured in zebra and in the Cisco's (by now, you should have figured out how to access the Cisco documentation on the Web).

Answer these questions in your lab report.

1. What commands did you use to configure RIP on the Cisco router? Include the running-config of the router. Call the file B.conf.
2. What commands did you use to configure ripd in zebra?

Setting up a Web server

The next step is to set up a web server on machines C, D and E of your topology. For this step, you will need to understand how to configure the Apache web server that comes with most Linux distributions. Basically, you will need to put the following two files into the right place in the file system (you may need to create a directory to do this):

• The XORP software router
• The XEmacs main page

so that, when you use a browser from machine A and type the following URL

http://ip_address_of_C/tcpexpt/xorp-1.0.tar.gz

the browser should start downloading the file.

You should install these two files on all three machines C, D and E since they will be needed for your experiment.

Answer the following questions in your lab report:

1. What commands did you use to configure httpd (the web server) to do this?
2. Instead of a browser, find a Linux command line tool that allows you to download web pages. Show me that your setup works by including in your submission a script of a Unix session where you actually download the two files on machine A (see the man page for the script command) using this command line tool.

To do this, you will write one or more Shell scripts that, when executed at A, will initiate three concurrent downloads of the large file (the XORP software), one from each of C, D and E (using the Unix command line tool for downloading web URLs that you found in the previous step). Your script will also concurrently (and repeatedly) download 3 copies of the smaller file from each of C, D and E. Thus, at any instant, you should try to (we say try to, because, while the short flows are stopping and starting again, there may be times when you have fewer than 9 short flows that are concurrently active) have 3 long flows and 9 short flows going on. The experiment stops when the last long flow is done downloading the large file.

While the experiment is on going, you should also collect binary packet traces on C, D, and E using tcpdump without filtering out any traffic. Read the man page for tcpdump carefully before embarking on the experiment.

Part A

Once the experiment is done, you should now be ready to analyze the results. To do this, you would use the tcptrace tool (see the Resources section above).

Answer the following questions in your lab report (or perform other actions as necessary):

1. Include the shell script along with your submission. Name the file tcp_fair.sh.
2. Compress the traces from each of C, D, and E and include them in your submission. These traces must be named C.tcp_fair.trace.gz, D.tcp_fair.trace.gz etc.
3. Include three summaries of these traces (i.e run tcptrace -b on them). These files must be text files and must be named: C.tcp_fair.summary, D.tcp_fair.summary etc.
4. Now use tcptrace to compute the average throughput given to a short flow. Similarly, compute the average throughput achieved by a long flow. What values do you get?
5. Are these two values different? Would you have expected them to be different? If the values turn out to be different, proceed to the next step. Otherwise stop here.
6. Include the time-sequence plot (see the tcptrace manual for this) for one short flow from machine C. Name the file C.tcp_fair.short.time-seq.eps. Include a time-sequence plot for the long flow from machine C. Name the file C.tcp_fair.long.time-seq.eps. Do these two files give you an idea why you saw a throughput difference in step 4? Explain.
7. Is TCP fair to short flows?

Part B

Now, keep the same overall configuration, but figure out how to configure RED (Random Early Detection) in router B (the Cisco router). Keep the default parameters for RED that Cisco uses.

Re-run the same experiments as discussed above and answer the following questions:

1. What commands did you use to enable RED on the Cisco?
2. Compress the traces from each of C, D, and E and include them in your submission. These traces must be name C.tcp_fair_red.trace.gz, D.tcp_fair_red.trace.gz etc.
3. Include three summaries of these traces (i.e run tcptrace -b on them). These files must be text files and must be named: C.tcp_fair_red.summary, D.tcp_fair_red.summary etc.
4. Now use tcptrace to compute the average throughput given to a short flow. Similarly, compute the average throughput achieved by a long flow. What values do you get?
5. Are these two values different? Would you have expected them to be different?
6. How do these values compare with the throughputs obtained using FIFO (the default queueing behavior)?

Now disable RED in the router so that it defaults to using FIFO queueing.

To do this, you will use ttcp or iperf to send UDP large (to determine the largest possible packet you can send, you need to understand what MTU means) packets from C to A. Read about ttcp and iperf and figure out how to send UDP traffic. Update your shell script to add this. That is, your script running on A should not only have the short and long flows as before, but it should also cause C to generate a sequence of back-to-back UDP packets.

Part A

Now, run the script, and obtain the tcpdump traces from machines C, D, and E, as before.

1. Include the shell script along with your submission. Name the file tcp_iso.sh.
2. Compress the traces from each of C, D, and E and include them in your submission. These traces must be named C.tcp_iso.trace.gz, D.tcp_iso.trace.gz etc.
3. Include three summaries of these traces (i.e run tcptrace -b on them). These files must be text files and must be named: C.tcp_iso.summary, D.tcp_iso.summary etc.
4. Now use tcptrace to compute the average throughput given to a short flow. Similarly, compute the average throughput achieved by a long flow. What values do you get?
5. Is each of these throughputs different from the corresponding throughput numbers obtained in Part A of the "Mice and Elephants" experiment? If they are explain what you think is happening, using one of the types of plots that tcptrace generates, for one of the flows (we don't tell you which plot will provide the most graphic explanation for what is happening; this is something for you to decide). Name this file iso_plot.eps.

Part B

Now re-enable RED on router B, and re-run the script above and collect the TCP traces.

Answer the following questions:

1. Compress the traces from each of C, D, and E and include them in your submission. These traces must be named C.tcp_iso_red.trace.gz, D.tcp_iso_red.trace.gz etc.
2. Include three summaries of these traces (i.e run tcptrace -b on them). These files must be text files and must be named: C.tcp_iso_red.summary, D.tcp_iso_red.summary etc.
3. Now use tcptrace to compute the average throughput given to a short flow. Similarly, compute the average throughput achieved by a long flow. What values do you get?
4. Is each of these throughputs different from the corresponding throughput numbers obtained in Part B of the "Mice and Elephants" experiment? If they are explain what you think is happening, using one of the types of plots that tcptrace generates, for one of the flows (we don't tell you which plot will provide the most graphic explanation for what is happening; this is something for you to decide). Name this file iso_red_plot.eps.

After doing this, disable RED in the router.

For this part, you will use the script that you used in the "Of Mice and Elephants" section.

The only change between this experiment and that one is that you will need to configure your network so that traffic from E to A gets higher priority. To do this, you will need to:

• Configure the router B to do weighted RED with two queues.
• Configure machine E to mark all packets to A with a certain value of the precedence field. We have given enough pointers for you to figure out how to do this.

Answer the following questions:

1. List the commands you used to configure the router B and machine E.
2. Compress the traces from each of C, D, and E and include them in your submission. These traces must be named C.tcp_diff.trace.gz, D.tcp_diff.trace.gz etc.
3. Include three summaries of these traces (i.e run tcptrace -b on them). These files must be text files and must be named: C.tcp_diff.summary, D.tcp_diff.summary etc.
4. Now use tcptrace to compute the average throughput given to a short flow through E (suppose we call this TS(E)). Similarly, compute the average throughput achieved by the long flow through E (TL(E)). Also compute TS(CD) (the average of short flows from C and D) and TL(CD). What are the values of these four variables?
5. Are TS(E) and TS(CD) the same or different? How do they compare with the average throughput of a short flow obtained in the first experiment (section "Of Mice and Elephants", Part A)?
6. Are TL(E) and TL(CD) the same or different? How do they compare with the average throughput of a long flow obtained in the first experiment (section "Of Mice and Elephants", Part A)?
7. If you found that TCP was unfair to short flows in Section "Of Mice and Elephants" part A, does this experiment suggest a solution to that fairness problem? What solution does this experiment suggest? What are the challenges in designing this solution?