Abstract
Performance analysis of applications on large clusters of SMPs requires a monitoring approach that supports tools realizing concepts like automation, distribution and on-line operations. Key goals are a minimization of the perturbation of the target application and flexibility and efficiency with respect to data pre-processing and filtering. To achieve these goals, our approach separates the monitor into a passive monitoring library linked to the application and an active 'runtime information producer' (RIP) which handles monitoring requests and performs pre-processing (e.g., aggregation) of performance data for individual cluster nodes. A directory service can be queried to discover which RIPs handle which nodes.
Original language | English |
---|---|
Pages (from-to) | 429-437 |
Number of pages | 9 |
Journal | Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) |
Volume | 2840 |
DOIs | |
State | Published - 2003 |