Detail

Dynamic Tuning of Distributed Applications on Computational Grids

Due recent increase of available computer resources, parallel applications may face heterogeneity in
different runtime scenarios considering network and processors properties. In these scenarios
developers cannot predict all possible parameters in order to get efficient resource use and lower
execution time given a set of resources. Dynamic changes of characteristics in runtime require dynamic
application performance monitoring, analysis and tuning. That approach is called dynamic tuning.
I will present an introduction of dynamic tuning and our Monitoring Analysis and Tuning Environment
(MATE) and its key concepts. In the dynamic tuning, the analysis is determined by performance models.
These models specify what should be measured during runtime and what can be changed in order to get
better performance indexes. We had applied dynamic tuning in cluster computing and our current
research is to apply same concepts to Computational Grids. Initially, we worked on process tracking and
application monitoring on these systems. We are currently evaluating performance model of
Hierarchical Master Worker applications to control the number of workers and application grain size.