K. Hayakawa, T. Sasaki, H. Umeda, and U. Nagashima (Japan)
Cluster Computing, Heterogeneous Computing, Network Interfaces, Molecular Orbital Calculations
The heterogeneous cluster systems allow us to get high per formance and low cost using new CPUs and existing CPUs. In case of adding new nodes to a cluster system, we can get higher performance CPUs than existing them with lower cost. In order to execute an application on the heteroge neous cluster systems efficiently, the task scheduler have to strike a balance between node’s performance and the amount of the tasks. The purpose of our work is to propose the best load balancing method(including the scheduler) of Molecular Orbital Calculations on the heterogeneous clus ter system. In the task scheduler, we have implemented static load balancing and dynamic load balancing meth ods. This paper shows 2 improvements to bump up paral lelization efficiency on the load balancing methods. One is scheduling multiple compute processes on static load bal ancing method, the other is improvement of network. As the result, EMDC achieved max 28.9% improvement of it in the multiple compute processes and also achieved 10% improvement of the parallelization efficiency in the net work improvement.
Important Links:
Go Back
