|
I'm a Phd student since April 2003. Supported by BULL-SA
(France), my works mainly consist in a large study of Input/Output characteristics from HPC Applications and existing parallel I/O systems.
Currently, I am studying the different behaviours of such systems, with a peculiar interest in the parameters ruling these from a hardware, middleware and software
point of view. The next step of my works will tackle the issues related to MPI I/O.
In addition since 2002, I have been working on the NFSP project which was launched by Pierre Lombard and Yves Denneulin in 2001. NFS
Parallèle is a distributed version of the ubiquitous Network File System. The main idea consists in exploiting the performance (bandwidth,
huge storage space, ...) coming from a beowulf cluster architecture.
|
Abstract :
I/O bottlenecks have always been a major issue in computer science.
The existing file systems try to exploit the growing capabilities provided by the recent architectures like clusters and the better hardware
(high bandwidth, computing ressources, huge storage space ...) Unfortunately, the complexity of such systems
leads to deal with several new drawbacks (heterogeneity, fault tolerant, scalability and security). Futhermore, lots of research groups attempt to develop
new kinds of file systems which are able to provide acceptable performances and an adequate API to large-scale applications (especially the ones that manipulate vaste datasets).
All these issues make I/O optimizations an important and challenging problem for cluster.
...
|
