| home | calendar | topics | work | resources |
|
|
CPS 214: Networks and Distributed Systems | |
|||||
|
|||||||
Our first two labs will be based on previous offerings of 214. The text below
on the first two assignments (Web Server and Transport Protocol) is based
the Spring 2004 offering of CPS 214. Varun has revised it to bring it
up to date for 2007. Please let us know of any problems!
All assignments should be done in teams of two. Please see me if you have any concerns regarding this. Also, you must choose a different person for each assignment, though the person you work with for the final project is your choice.
For the checkpoints, I just want you to send an e-mail with checkpoint in the title, tell me who you will be working with (if anyone), a brief description of your plan (e.g., what tools will you use and how long do you think the various pieces will take), a description of any initial progress, any questions you might have. Also, please let me know if you think there are resources you might need to get it done that you do not already have.
January 25 checkpoint, due February 8.
Grade criteria
Note:
Several questions have come up regarding the lab and use of software
components
that you did not write yourself. As a matter of principle, it is not my
intent
to cause you to waste your time on busy work, although busy work is part of
software development and a part of life. I refer you to the statement of
course policies. The guideline is admittedly
vague, but here is a rule of thumb: don't use any components that you could
not implement easily yourself. If you use a component from somewhere else,
you should be able to explain in detail what it does and how it works.
The goal of this project is to build a functional HTTP/1.0 server. This assignment will teach you the basics of distributed programming, client/server structures, and issues in building high performance servers. While the course lectures will focus on the concepts that enable network communication, it is also important to understand the structure of systems that make use of the global Internet.
At a high level, a web server listens for connections on a socket (bound to a specific port on a host machine). Clients connect to this socket and use a simple text-based protocol to retrieve files from the server. For example, you might try the following command from a UNIX machine:
% telnet www.cs.duke.edu 80
GET / HTTP/1.0\n
\n
(type two carriage returns after the "GET" command). This will return to you (on the command line) the html representing the "front page" of the Duke computer science web page.
One of the key things to keep in mind in building your web server is that the server is translating relative filenames (such as index.html) to absolute filenames in a local filesystem. For example, you might decide to keep all the files for your server in ~student/cs214/server/files/, which we call the root. When your server gets a request for /index.html, it will prepend the root to the specified file and determine if the file exists, and if the proper permissions are set on the file (typically the file has to be world readable). If the file does not exist, a file not found error is returned. If a file is present but the proper permissions are not set, a permission denied error is returned. Otherwise, an HTTP OK message is returned along with the contents of a file. How would your server support ".htaccess" files on a per-directory basis to limit the domains that are allowed access to a given directory?
You should also note that web servers typically translate "GET /" to "GET /index.html". That is, index.html is assumed to be the filename if no explicit filename is present. The default filename can also be overridden and defined to be some other file in most web servers.
When you type a URL into a web browser, it will retrieve the contents of the file. If the file is of type text/html, it will parse the html for embedded links (such as images) and then make separate connections to the web server to retrieve the embedded files. If a web page contains 4 images, a total of five separate connections will be made to the web server to retrieve the html and the four image files. Note that the previous discussion assumes the HTTP/1.0 protocol which is what you will be supporting in this first assignment.
Next, add simple HTTP/1.1 support to your web server, consisting of persistent connections and pipelining of client requests to your web browser. You will also need to add some heuristic to your web server to determine when it will close a "persistent" connection. That is, after the results of a single request are returned (e.g., index.html), the server should by default leave the connection open for some period of time, allowing the client to reuse that connection to make subsequent requests. This timeout needs to be configured in the server and ideally should be dynamic based on the number of other active connections the server is currently supporting. That is, if the server is idle, it can afford to leave the connection open for a relatively long period of time. If the server is busy, it may not be able to afford to have an idle connection sitting around (consuming kernel/thread resources) for very long.
For this assignment, you will need to support enough of the HTTP protocol to allow an existing web browser (Firefox, Safari or Konqueror) to connect to your web server and retrieve the contents of a sample page from your server. (Of course, this will require that you copy the appropriate files to your server's document directory.)
At a high level, your web server will be structured something like the following:
Forever loop:
Listen for connections
Accept new connection from incoming client
Parse HTTP/1.0 request
Ensure well-formed request (return error otherwise)
Determine if target file exists and if permissions are set
properly (return error otherwise)
Transmit contents of file to connect (by performing reads on
the file and writes on the socket)
Close the connection
You will have three main choices in how you structure your web server in the context of the above simple structure:
You may choose from C,C++, or Java to build your web server. In C/C++, you will
want to become familiar with the interactions of the following system calls to
build your system: socket(), select(), listen(), accept(), connect().
We outline a number of resources below with additional information on these
system calls. A good
book is also available on this topic.
If you choose to implement your web server in Java, you may find
pre-written classes that implement http for you. Be warned that modifying
those classes is likely to be much more work than implementing the server
yourself. If in doubt about whether to use a particular class,
please ask first. In general, the only classes you will need will come with
Sun's JDK.
Now that you have a functional web server, the second part of the project involves evaluating the performance of the system that you have built. Build a synthetic client program that connects to your server, retrieves a file in its entirety, and disconnects. The goal of this load generator is to evaluate the performance of your server under various levels of offered load. You will measure server performance in terms of throughput (requests/sec) and in terms of latency (average time to retrieve a file). Your synthetic load generator might be multi-threaded, with a different number of threads attempting to retrieve files as quickly as possible, or you may use a multi-process approach, where individual instances of your load generator start up, retrieve a file, and shut down. You can control offered load by controlling the number of simultaneous threads/processes that are retrieving files from the server. You should measure how long each request takes and keep track of this information for summary reporting (this will be used to generate average latency). You should also measure the total number of requests satisfied in a given time period (this will be used to generate average throughput).
A principal aspect of this assignment is to compare the performance of your web server using HTTP/1.0 versus HTTP/1.1, especially given the behavior of TCP. There is significant overhead to establishing and tearing down TCP connections (though this is less noticeable in a LAN setting) and persistent connections avoids this issue to some extent.
Can you think of any instances where 1.1 will underperform 1.0? A number of considerations affect the performance HTTP/1.0 versus 1.1. Principally consider some of the pros and cons of using a connection per session versus using a connection per object. Principally, the difference between the two comes down to the following:
Under what conditions would HTTP/1.1 underperform 1.0? Also consider the effects of the bandwidth delay product, round trip times, and file sizes on this tradeoff. For example, high round trip times exacerbate the negative effects of slow start (taking multiple rounds to send a file even if the bottleneck bandwidth would allow the entire file contents to be sent in a single round trip).
It will be important to run your web server and load generator on different machines. It may be necessary to use multiple machines as load generators to saturate the server. You will want to keep track of CPU load on both the load generator and the server machines to determine when each system component saturates (at least with respect to CPU load).
Profile your server code to get a rough idea of the relative cost of different aspects of server performance. For different sized files, how much time is spent in establishing/tearing down connections versus transmitting files? Use timing calls around important portions of your code to keep track of this time on a per-logical operation basis. What type of bandwidth does the web server deliver as a function of file size?
Also try running your test for different size files. How does latency/throughput change with file size?
You may find it useful to use a scripting language, such as perl, to control the performance evaluation.
For extra credit, conduct your performance evaluation using different structures for your HTTP server. For example, if you decided to build your server using a multi-threaded model, implement an alternate version that uses an event-based/single threaded approach for handling requests. How does the performance characteristics of your server change with different system structure?
Writeup
A very important aspect of your assignment will be your project report describing your system architecture, implementation, and high level design decisions. For your performance evaluation, you should include a number of graphs describing the various aspects of system performance outlined above. Make sure to clearly describe how you set up your experiments. What kinds of machines did you run on? What kind of network interconnected the machines? How did you build your load generator and collect statistics? How many times did you run each experiment to ensure statistical significance? Ideally, you will include error bars to indicate standard deviation or 95% confidence levels. Finally, your writeup should include explicit instructions on how to install, compile, and execute your code.
You should turn in an initial version of your writeup for the first part of the assignment and then add to it (with any necessary revisions) for the second part. Your writeups are due with the assignments at the times specified above.
Submission:
Submit your files (c, c++ or java) with compiling instructions via email to
the TA at {varun at cs.duke.edu}. You should submit both your server and
the client you used to test your server. The final version of the writeup
should also be submitted. Please do not submit compiled code, only the
sources.
If you are using java, please use the most recent version. If you are
compiling on the acpub machines, a slightly newer version of java is in
/afs/acpub/project/cps/bin/java.
Resources
We have collected a list of available resources to help you get started with various aspects of this assignment:
Due March 1.
Grade criteria
The goal of this project is to build a simplified version of TCP on top of the UDP protocol. You are responsible for defining your own header format, which does not have to be compatible with any TCP implementation. The protocol should support:
The code should be written in C/C++ as a multi-threaded program. You do not need to reproduce the Sockets API, but you should, at a minimum, provide send() and recv() functions that your test application calls. You will need to write both the sending side and the receiving side (full-duplex is required, meaning that your sender and receiver test applications could call both send() and recv() functions). On each side there will be a thread that attempts to read or write an infinite data source, representing the user-level program. A second thread manages packet transmission and reception, retransmission times, advertised windows, congestion windows, sequence numbers, etc.
You will need to implement the equivalent of a producer/consumer queue of some finite length. On the sending side, if the sending thread attempts to write to a queue that it is full, the thread should block (be put to sleep) until room is available in the queue. On the receiving side, if the receiving thread attempts to read from a queue that is empty, it should block until (in order) data is available.
The sender should wait until it has a full packet's worth of data buffered before transmitting (assume a transmission size of 1000 bytes). Of course, the sender should only send up to its window size, which will be the minimum of the advertised window from the receiver and the slow start rate.
Managing retransmission timers will be a little bit tricky. Options include periodically checking if any set timers have expired on the sender side. or spawning threads that sleep until a certain amount of time has passed (e.g., by using the select system call) and then signalling the main sending thread that a packet needs to be retransmitted. The first approach is less efficient while the second approach is potentially more complicated.
You should use Van Jacobson's formula to estimate round trip time, accounting for both the exponential weighted moving average and the variance. You will need a data structure to timestamp each packet transmission and to read the system clock on the reception of each acknowledgment.
You do not need to implement fast retransmission, fast restart, or additive increase/multiplicative decrease. However, teams doing so will receive extra credit. You do however, need to implement ack pacing, with each acknowledgment allowing the injection of a new packet into the network.
Writeup
You should include a detailed writeup of your architecture along with your project submission. How did you structure your code? What are your packet formats? What is the high level flow of events through various threads in your code?
Submit
Submit your files (c, c++ or java) with compiling instructions via email to the TA at {varun at cs.duke.edu}. The final version of the writeup should also be submitted.