Tackling the complexity of Grid programming with Self Adapating Numerical SoftwareHeidelberg 31 May 2002 Writing efficient numerical software for a complex, ever changing environment as the Grid, can be a nightmare at worst, and a very time consuming activity in the best case. Self Adapating Numerical Software (SANS) could be a solution, explained Jack Dongarra at the Heidelberg ISC2002 conference.
Today, there seems to be a tendency to focus on hardware. Hence the gap between usable and deliverable performance is increasing. You get performance out of a system only if the data and controls are setup just right. Otherwise there are dramatic performance degradations. This poses challenges for libraries, PSEs and tools, that have tor each Tflop/s or later Pflop/s perfomance.
There are all kinds of software issues that have to be addressed. For numerical software predictability of accuracy and performance are key. In dynamic environments, run-time selection of algorithms and resource management to make the best use of the available hardware is also important. The Grid also requires algorithms that are tolerant for large and unpredicatable latencies.
To make life easier for users of numerical software, a SANS effort (Self Adapating Numerical Software) has been started. This is software that obtains information on the underlying system where they are running on. During execturion optimization and pehaps reconfiguration based on newly available resources must be possible.
Optimizing software to exploit features of a new processor has historically been an exercise in hand customization, said Dongarra. It is a time consuming and tedious process that has to be done over and over again for very new architecture. Also the number of computational kernels that have to be optimized is growing. So there is a need for quick and dynamic deployment of optimized routines.
SANS will provide software technology to aid in reaching high-performance on commodity processors, clusters, and Grids. It will consist of a pre-run time and a run time optimization. Tools for performance modelling and analysis will be built in. SANS will automatically select the best algorithm. Dongarra used the ATLAS Blas software, as an example where automatic tuning work has been done.
A lot of information on numerical libraries and software can be found at: Netlib.
Ad Emmen [News on Advanced IT] [Calendar] [Analysis] [IT in Medicine] |
 |