Knowledge-based Performance Monitoring for Large Scale HPC Architectures
Guillén Carías, Carla Beatriz · 2015
This thesis presents concepts for systemwide monitoring and performance analysis of HPC systems, which are aimed at a preliminary detection of inefficient applications. On-line analyses without instrumentation of user codes are performed with codified expert knowledge designed to reveal bottlenecks in running applications. Novel optimizations to collect data and data reduction techniques ensure an efficient and scalable monitoring of HPC architectures. Detailed results are provided for a petaflop system.