In an application, data access can become a major performance bottleneck if the memory hierarchy of the underlying hardware architecture is not taken into account. The only way to gain deeper insight of an applications memory usage is to measure its data access behavior with hardware counters. From the programmer’s point of view such performance data (like cache misses or hits) have to be linked to the data structure causing it. The name of a data structure is the only point of reference the user has and the only point where he can apply optimizations. In the project EP-Cache tools for Fortran applications were developed to monitor and to link hardware counter information with data structures. This includes an appropriate visualization of the gathered information. Parallelism using the OpenMP programming paradigm is also supported.
|Seiten (von - bis)
|NIC series : publication series of the John von Neumann Institute for Computing (NIC)
|Veröffentlicht - 1 Okt. 2005