1st International ICST Conference on Scalable Information Systems

Research Article

Network system design affects distributed parallel computing

  • @INPROCEEDINGS{10.1145/1146847.1146891,
        author={Jin Guojun and Frank Wang},
        title={Network system design affects distributed parallel computing},
        proceedings={1st International ICST Conference on Scalable Information Systems},
        publisher={ACM},
        proceedings_a={INFOSCALE},
        year={2006},
        month={6},
        keywords={Jumbo Frame Network on Chip (NoC) System Resources System on Chip (SoC) Distributed Computation network Design TCP},
        doi={10.1145/1146847.1146891}
    }
    
  • Jin Guojun
    Frank Wang
    Year: 2006
    Network system design affects distributed parallel computing
    INFOSCALE
    ACM
    DOI: 10.1145/1146847.1146891
Jin Guojun1,*, Frank Wang2,*
  • 1: Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, CA 94720
  • 2: Cambridge-Cranfield High Performance Computing Facility, United Kingdom
*Contact email: jin@george.lbl.gov, f.wang@Cranfield.ac.uk

Abstract

Computational Grid is essential for scientific advance, and advanced technology research and development because single system computing power is restricted by processor clock speed advancing. Increment of processor clock speed starts decreasing due to the minimum size of silicon dies v.s the distance that electrons can travel in one clock cycle driven by the light speed. The maximum processor clock speed affects not only computation power, but also communication capability because current network subsystem relies on the system resources, such as CPU, to move data from/to network. To provide a high performance and scalable network subsystem for distributed computation systems, intuitively, making network subsystem to use less or not to depend on these system resources is the key. This paper will analyze system resources consumed in each network subsystem layer/module, and address what design decision in network subsystem advance, such as Jumbo Frame, Network on Chip (NoC) --- to remove system resource consumptions for network operations --- will be able to increase the computation power on distributed computational systems.