Efficient Management of Huge Data Sets in Cluster Computers

                       Hipolito Vasquez

Conveying data distributed among different compute nodes to a file
striped across storage devices, in order to exploit the inherently
aggregate bandwidth, is the basic characteristic of parallel file
systems. Because applications have different I/O requests, it is
unlikely to obtain high performance through the use of one single
distribution policy. This work proposes an adaptive on the fly
distribution framework for parallel file systems. In order to apply
the most suited distribution policy at a certain point in time,
information will be acquired from an adaptive file access pattern
recognition mechanism.