Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
488177 | Procedia Computer Science | 2011 | 8 Pages |
In high performance computing, many large scientific and engineering problems are solved on a supercomputer which is the gathering of two specialized entities, one dedicated to computations and another one to I/O. Many applications to settle problems have deterministic behaviors in computations and I/O. This knowledge may be used to load data in advance or delegate data writing on dedicated nodes. Thereby, it could be interesting to use the specialized parts of the supercomputer and this knowledge in order to have a better cache management by uncoupling computations and I/O. This has led to the design and evaluation of a first prototype of ASIODS. This paper presents the architecture of our approach and the results obtained showing the concept capabilities. We demonstrate that the approach reduces the execution time by avoiding I/O access penalties.