EP-Cache

Overview

The performance gap between high-performance microprocessors and memory is still increasing. Especially for modern parallel systems this is gaining more and more importance since the basic building blocks are SMP nodes, i.e. multiple processors access the same physical memory.

The goal of the EP-Cache project is to develop new performance analysis tools and performance tuning techniques for programming SMP nodes effectively. These techniques will be based on new hardware monitoring support which will provide much more information on the access behavior of applications than the current performance counters integrated on the CPUs. The monitors will gather information on cache misses and coherence actions for individual data structures and code regions. The performance analysis tools will provide visualization support as well as automatic analysis support based on the results of the APART working group. For performance tuning, program transformations from HPF compilers will be investigated as well as standard transformation for locality optimization. Programming tools will support the user in applying the transformation to the given code.

The project is funded for three years by the German Federal Ministry for Education and Research until February 2005.

Our tasks

bullet

Develop designs for new hardware performance monitors including an estimate for the realization costs.

bullet

Implementation of an evaluation environment that simulates the hardware monitors in the context of OpenMP programs on Intel-based SMP-nodes.

bullet

Develop an automatic control of the hardware monitor to gather appropriate information taking into account the limited resources (e.g. counters) in the hardware monitor.

Partners

bullet

FhG/SCAI: Institute for Algorithms and Scientific Computing (Coordinator)

bullet

Pallas GmbH

bullet

Technische Universität Dresden, ZHR: Center for High Performance Computing

bullet

Technische Universität München, LRR: Lehrstuhl für Rechnertechnik und Rechnerorganisation / Parallelrechnerarchitektur

Contact

Prof. Dr. Michael Gerndt
Technische Universität München, LRR
80290 München
Dr. Wolfgang Karl
Technische Universität München, LRR
80290 München

Publications