Ohm Performance Graph Applet

One of the primary goals of Calvin's Ohm project was to determine the performance and cost-effectiveness of the beowulf cluster in star+ring and star+hypercube network toplogies. Because these topologies require extra network hardware, they are more costly than the traditional star topology. The applet below displays the results of our tests for many different software applications:

Applications were divided into three groups based on the amount of network traffic generated by the application. Those with the highest amounts of network traffic were considered 'fine grained' applications, those with the lowest amounts were considered 'coarse grained' and the applications in between were considered 'medium grained.'

Each application was then run many times on the beowulf cluster for each configuration:

• single processor
• 16 node star topology
• 16 node star+ring topology
• 16 node star+hypercube topology

The speedup of an application under a particular configuration was determined by dividing the time required to run on a single processor by the time required to run on the particular configuration. The cost of the network was determined using the costs of our switch and network cards. For more information of Ohm's hardware configuration, please see here.

The result that is actually seen in the plots above is the speedup for a particular topology divided by the price of that topology. Thus, larger numbers indicate higher performance per dollar spent on the network. By changing the price of the switch and the network cards purchased, it may be the case that as network hardware costs drop, the cost-effectiveness of the star+ring and star+hypercube topologies will rise. When the Ohm beowulf cluster was originally constructed, the hardware costs were too high to render the additional network hardware cost-effective. However, at current prices the star+ring and star+hypercube topologies are more cost-effective than the star topology for some parallel applications.