The hpcLine cluster at CEASaclay 23 May 2003 The French energy research centre CEA has a 48-processor hpcLine installed for a CFD project for nuclear applications. At the SAVE meeting in Saclay in May, Christophe Calvin from CEA talked about the experience they had with the machine.
The code CEA is running on the hpcLine is Trio_U. This is a CFD code for nuclear applications that can be applied to incompressible flows in large and complex geometries. The scales for which Trio_U are used, range from micro to mesa scales.
Trio_U is and Objected Oriented code, programmed in C++. The parallel parts use MPI. This approach leads to a code that is portable from PC to supercomputer. On industrial applications, the efficiency reached is 75% to 90%.
Why use a cluster like the hpcLine, if the code can run on a PC and on a supercomputer? CEA develops on PCs, said Christophe, large parallel production is done on a HP SC232, but small, i.e. less than ten processors parallel production and sequential production fit well
the hpcLine cluster.
The 48 CPUs of the hpcLine at CEA have different speeds, ranging from 1.7 GHz to 2.2 GHz and 1 Gbyte of memory and Scali SCI interconnect. The machine runs Linux Redhat with SSP SCALIUNIVERSE, OpenPBS, PGI, LAM/mpi, mpich, SCAMPI, and LAPACK.
The machine runs reliable with a 60% utilisation since its installation and is easy to manage. Only the file storage is not satisfactory yet. Users migrate their data and results to their own
PC. In the future, CEA would like to go to an environment were there is a large common disk space with automatic data transfer between PC and cluster when needed.
Ad Emmen |
 |