The Effects of Cluster Size on Packet Inter-arrival Patterns and Intensity in a Distributed System

Although the design and implementation of parallel processing infrastructure is not considered a mainstream IS topic, a working knowledge of its capabilities can have an important influence on improving the design of on-line systems. In order to illustrate some of the advantages and limitations of using parallel processing in information systems two experiments were run to collect multiple node scaling information. Specifically, packet inter-arrival data was collected and analyzed for a message passing interface (MPI) and an HTTP problem that simulated packet movement in an enterprise level LAN test-bed. The parallelization method that we employed broke the problem into N subparts based in the number of processors used. Inter-processor communication was required whenever a processor needed to interact with another one to help solve the application. That communication was relayed via the MPI protocol using TCP packets with intensities in .0005 second range. Experimental trials were run on various platforms: from 2 to 12 two processor units for MPI, and 8 and 16 units for HTTP. In the MPI experiment, although the CPU time continued to drop as additional units were added, the elapsed time only dropped to the 4 unit level and then increased thereafter. In the HTTP experiments the two switch/server model provided only a slightly better performance than the one switch/server model.