CS/ECE 757 Fall 2000

Homework 2: Beowulf Bakeoff

Due: In class Tuesday 11/21/2000

Last Updated: 11/15/00

For the Beowulf Bakeoff, the class will break into three groups of four people. Each group will build a cluster made up of six nodes and each node will contain two processors. At the end of the project, we will connect the three clusters together to form a 20 node, 40 processor cluster. We will pick the cluster to use as the base for the large cluster based on the features that the group implemented, performance, scalability, maintainability, etc. The group that "wins" (i.e., the group that builds the cluster that we choose to use as the base of the large cluster) gets the six Celeron processors that we started building with as a prize.

Hand in:

Below is a list of things that need to be handed in:

List of implemented features and description of features
User's guide of how to use features of cluster, how to scale cluster (i.e., how to add nodes), etc.
Graphs of performance of cluster

Features:

A minimal cluster must have the following features:

A master node with a disk and five diskless slave nodes
MPI
Use to ECE's Yellow Pages server
NFS mount ECE home directories

You may also wish to install the X libraries, since you will need this for the programming assignment (homework 3).

Performance Analysis:

For the project, you must run the following performance tests and produce graphs for each. These performance tests are from MPI (if your version of MPI did not come with these test programs, you can get them by downloading MPICH from the MPI web page. These tests are in "mpich/examples/perftest". Note that results from these tests generally vary each time they are run, so you should run each test multiple times and report the average results.

Run a stress test on the cluster using the following command:
This measures the bandwidth of the system by sending a large message (size from 400K to 800K, increasing by 12.5K each step) from each processor in turn to every other processor. Vary the number of processors from 2-12 and graph the average aggregate bandwidth for each case.
Run the following test on the cluster using the following command:
This measures the speed of each communication channel in the system by having each processor send a small token (on the order of a word or two) to its neighbors. By doing this one both gets a measurement of the overhead of sending messages and also see if different links are different speeds. If all of links are about the same speed, tcomm will simply report the average bandwidth. However, if there are large differences in the speed, tcomm will print out a histogram of the speeds. Vary the number of processors and graph the average bandwidth for each case. In the cases where tcomm finds that there is a large difference in speed, report the differences that tcomm finds.

Feel free to also hand in graphs for any other performance tests that you perform, as these will help us choose which cluster to use as the base. You may wish to perform many of these, since they will give you a better idea of the behavior and performance of your cluster, which will give you some insight into how to write the parallel program for the next assignment.