|
|
||||
|
OPNET Technologies
OPNET is a registered +++
Department of Computer Science
Sponsoring Professor: Charles Nicholas Graduate Student: Frances Roth |
Simulation of a Beowulf Style ClusterIn recent years, cluster computing has come to dominate the area of high performance computing. This has brought with it some problems not found in earlier supercomputers. One problem is the lack of shared memory. Without shared memory, it became the programmer's responsibility to ensure that data was transferred between cluster nodes as needed by making explicit calls to a library such as MPI. Attempts have been made to get around this problem through the use of Distributed Shared Memory (DSM) or programming languages which build implicit data transfers into the language itself. These solutions free the programmer from the need to know exactly which data resides on which node. They do not, however, make the data equally accessible to all nodes. We believe that by placing additional memory at the central switch in a Beowulf style cluster, we can create a shared memory which is equally accessible to all nodes in the cluster and, furthermore, improve the access times for most data which is not local to the requesting node. In an effort to demonstrate this, we are using the OPNET Modeler program to build a simulation of a Beowulf cluster. The first version of the model will be used to provide estimates for the performance of various memory access patterns on a generic Beowulf cluster when MPI routines are used. These data can be used to verify the basic model by comparing the model results with those obtained by running programs with similar memory access patterns on an actual cluster. The second version of this model will include a RAMSwitch at its center. The RAMSwitch is a modification of a standard switch with additional features that make it, among other things, easier to implement shared memory at the switch. This version of the model will be used to classify the types of programs which can be expected to provide better performance when shared memory at the switch is used. |
|||