| Autor: |
Molchanov, I. N., Galba, E. F., Stepanets, N. I. |
| Zdroj: |
Cybernetics and Systems Analysis; July 1983, Vol. 19 Issue: 4 p456-464, 9p |
| Abstrakt: |
Conclusions The analysis of parallel realizations of iterative methods for a specific MPC architecture leads to the following conclusions.1.The efficiency of the parallel algorithms is largely dependent on the communication speed in the system. This may prove to be the decisive factor for Ke. It is particularly significant in problems where the parallel algorithm accesses the external memory of the MPC.2.The parallelism efficiency decreases as the AP speed increases (the MPC parameters remaining constant).3.If the problem can be solved using only the combined AP memory, Ke decreases with the increase in the number of APs (for a problem of given size) since the number of communications and synchronizations increases.4.For a given problem size and fixed MPC parameters, the maximal Ke is attained when information is stored and processed in APs with sufficient combined memory. If the required information does not fit into the combined AP memory, second-level storage has to be used, which reduces Ke and increases the problem-solving time. When the number of APs is more than necessary, the parallelism efficiency diminishes.5.When new methods or new algorithms are designed on the basis of existing methods for the given MPC architecture, it is essential to ensure that the parallel processing involves the minimum possible number of communications. It is advisable to reduce the number of communications by increasing the load of arithmetic operations on each AP. |
| Databáze: |
Supplemental Index |
| Externí odkaz: |
|
Full text from SpringerLink