Sumix is a cluster of PCs, i.e. a group of 21 computers connected over a special network and dedicated to scientific calculations. It can be used like a single big computer with special parallels programs made with MPI or like individuals computers running tasks with a load balancing system.
It's an heterogeneous cluster made with 10 HP X4000 workstations (with 2*2GHz P4 processors per nodes) and 10 DELL Poweredge 1750 (with 2*3GHz P4 processors) for a total amount of memory of 40Go. The 10 HP computers are connected with a very fast low latency SCI network. Nodes are running the Linux operating system and we use OpenPBS as load balancing system. A special node is dedicated to the adminitrations tasks (compilation, NFS server, PBS server)

The SCI network is based on a 2D topology like on diagram:

With the SCI network, the sustained data bandwidth exceeds 320 MBytes/s for a latency less than 4 usecs. Parallels developments are based on scaMPI, a special version of the Message Passing Interface which bypass traditional TCP/IP network. It is possible to launch MPIch job with up to 40 processors.

A detail of the SCI network

For more informations, please contact: Sylvain Ferrand