Solving non-smooth unconstrained optimization problem with LAMGAC in a LAN-WLAN grid domain

Abstract

Traditionally, the Local Area Network (LAN) has been used for parallel programming with PVM and MPI. The improvement of communications in Wireless Local Area Networks (WLANs) achieving till 11 Mbps make them, according to some authors, candidates to be used as a resource for Grid Computing. In this paper we use our library based on LAM/MPI named LAMGAC in order to parallelize an algorithm that finds the global minimum of a nonlinear real valued continuous function. The algorithm uses a strategy based on the division of the domain into small boxes and it locates the extreme by means of a multiple start algorithm (MRS). The local minimizer is carried out by means of the steepest descent and the DFP method. The novel of this approach is that we can vary, the parallel virtual machine in runtime (spawning new processes using functions defined in MPI-2), we generate algorithms in which computations and communications are efficiently overlapped and we include a web interface to offer our system as a Grid resource. At this moment we have measured the execution time of some algorithms and the components of LAMGAC, obtaining interesting results.