Before the nineties, I/O performance growth has not kept up to pair with progress in other subsystems, e.g. CPU and network I/O. As a consequence, bottlenecks observed in the execution of demanding applications have progressively shifted from these areas to the I/O subsystem, and performance gains were lower and lower.<P>
The solutions adopted in the last decade for fast I/O in supercomputer and MPP architectures improved substantially the performance of those applications, but are inadequate for modern architectures based on low cost COTS (Common off-the-shelf) components.<P>
This presentation will address "state of the art" components that enable us to build high performance computacional architectures (including SMP, cc-NUMA, MPP and clusters) but main focus will be on I/O components: disks, disk arrays and interconnection topologies between servers and storage.<P>
We will conclude our presentation by focusing on the issues that have to be tackled by File Systems running on either shared nothing or shared disk cluster architectures.
