Abstract |
---|
The performance of a parallel system with NUMA characteristics dependson the efficient use of local memory accesses. Programmingand tool environments for such DSM systems should enable andexploit data locality. In this paper we present an event-driven hybrid monitoring conceptfor the SMiLE SCI-based PC cluster. The central partof the hardware monitor consists of a content-addressable counterarray managing a small working set of the most recently referencedmemory regions. We show that this approach allows to provide thenecessay and detailed run-time information which can be exploited byperformance evaluation and debugging tools. |
Contact |
Wolfgang Karl Lehr- und Forschungseinheit Informatik X,Rechnertechnik und Rechnerorganisation (LRR-TUM),Institut f{\"u}r Informatik, TUM,D-80290 Munich, Germany, karlw@informatik.tu-muenchen.de |