Analysis of cluster compute nodes with varying memory hierarchy distributions

This thesis compares a Compute Node, a cluster compute node that can completely contain smaller processes within, but not larger ones, with a Memory Node, a cluster compute node that can fully contain large processes within main memory. The Compute Node and Memory Node are tested by executing matrix multiplication programs that range in size from small processes similar to jobs that can be run on a single workstation to large processes that would only fit entirely within a Memory Node. Memory was allocated using four types of memory allocation: static, external, dynamic, and automatic. This analysis provides a measure of the benefit obtained by using a Memory Node over a Compute Node for large processes as well as a measure of the penalties that add up when a process is forced to utilize the swap device continuously during execution. The test programs executed showed the benefit of using a Memory Node over a Compute Node, with speedup values peaking at approximately 700 for the process sizes used. Additionally, the tests showed that when a Compute Node is forced to swap to the hard disk multiple times during execution, the majority of the execution time will be devoted to accessing the hard disk, not performing computations. Furthermore, the results showed that the method of allocating memory does not make a significant difference in execution times