Formulating the real cost of DSM-inherent dependent parameters in HPC clusters
Mohsen Sharifi, Ehsan Mousavi Khaneghah, Alfredo Tirado‐Ramos, Seyedeh Leili Mirtaheri · 2010
The choice of an appropriate interprocess communication (IPC) mechanism is critical to the performance of distributed systems and parallel programs. There is however a trade-off between the accrued system performance and the imposed cost of the deployed IPC mechanism. Message passing (MP) and distributed shared memory (DSM) mechanisms have been extensively studied and compared, but the state of the art work on formulating the cost of DSM is not comprehensive. In this paper we distinguish between DSM-inherent and application-specific parameters that contribute most to the real cost of DSM, and focus on DSM-inherent parameters in high performance computing (HPC) clusters. The weights of each parameter's effectiveness on the cost, as well as the weights of their influence on each other are determined. The derived formula for the real cost of DSM is then calibrated by a clustering coefficient that varies with different sizes of clusters. We have used the principles of management and accounting sciences in calculating the real cost of DSM.