I "cluster" di computer, anche chiamati Beowulf Computer sono una delle tecniche oggi una delle tipologie più usate nella realizzazione di supercomputer. Si basano sulla connessione in rete di computer che si suddividono il peso dell' esecuzione dei calcoli necessari

Il Progetto Hydro

In collaborazione con Istituto di Idraulica dell'Università Politecnica delle Marche è stato realizzato un cluster di computer per la parallelizzazione del software NLSWE di analisi del moto ondoso.

Hydro

HYDRO is being built on the basis of a cluster of 5 Intel (1.2 Ghz) Workstations with 256 MB of RAM and a 20 GB Quantum hard disk.. Each Workstation has one on-board one on-board Fast Ethernet card while 16-port Fast Ethernet switches have been used to connect the Workstations. The operating system is Linux (Suse Linux 7.3) and parallel computing is done via the LAM implementation of the MPI standard and High Performance FORTRAN. The Cluster is connected to the Istituto di Idraulica network by a second Fast Ethernet card. The master node is accessed by remote connection (telnet, rlogin ecc..) for software development, using the client/sever capability of Xwindows, and for running the different types of jobs required. We plan to use this computer for running predominantly the wave-resolving Nonlinear Shallow Water Solver.

Il modello Parallelo

• The parallel programme uses the MPI paradigm implemented in the LAM library.
• The code is a Single Program Multiple Data, using an approach Master/Worker.
• The master cluster node reads the input data, decomposes the full domain, generates the pointers, distributes the data to the workers, gathers the results from the workers, and prints the output.
• Each process has the same memory model but each worker acts only on a subset of the full memory through the use of suitable system of pointers generated by the master. In this way it has been easly port all the routines written for the sequential code to the parallel solver.
• Derived data type has been used which allows the sending/receiving of subset of matrices of data with only one operation.

Riferimenti

Riferimenti più approfonditi possono essere trovati nel seguente articolo

-- RobertoBernetti - 08 Jun 2005