Tesis doctorales: TESEO

Ayuda Salir

Ficha de Tesis

Imprimir

Ministerio de Universidades - Gobierno de España

Título: DYNAMIC TUNING FOR LARGE-SCALE PARALLEL APPLICATIONS
Autor: Martínez Trujillo, Andrea
Universidad: Universidad Autónoma de Barcelona
Departamento: ARQUITECTURA DE COMPUTADORES Y SISTEMAS OPERATIVOS
Fecha de Lectura: 10/09/2013
Mención Europeo / Mención Internacional: Concedido
Programa de doctorado: Computación de altas prestaciones
Dirección:
- SORRIBES GOMIS, JOAN (DIRECTOR)
- SIKORA, ANNA BARBARA (DIRECTOR)
Tribunal:
- Cunha, José (PRESIDENTE)
- Cortés Fité, Ana (SECRETARIO)
- ORTEGA LOPERA, JULIO (VOCAL)
Descriptores:
- ORDENADORES DIGITALES
Localización: BIBLIOTECA DE COMUNICACIÓN Y HEMEROTECA GENERAL UAB

2013martidynam.pdf

Marcador: https://www.educacion.gob.es/teseo/mostrarRef.do?ref=1044531Añadir a favoritos.
Resumen:
The current large-scale computing era is characterised by parallel applications running on many thousands of cores. However, the performance obtained when executing these applications is not always what it is expected. Dynamic tuning is a powerful technique which can be used to reduce the gap between real and expected performance of parallel applications. Currently, the majority of the approaches that offer dynamic tuning follow a centralised scheme, where a single analysis module, responsible for controlling the entire parallel application, can become a bottleneck in large-scale contexts. The main contribution of this thesis is a novel model that enables decentralised dynamic tuning of large-scale parallel applications. Application decomposition and an abstraction mechanism are the two key concepts which support this model. The decomposition allows a parallel application to be divided into disjoint subsets of tasks which are analysed and tuned separately. Meanwhile, the abstraction mechanism permits these subsets to be viewed as a single virtual application so that global performance improvements can be achieved. A hierarchical tuning network of distributed analysis modules fits the design of this model. The topology of this tuning network can be configured to accommodate the size of the parallel application and the complexity of the tuning strategy being employed. It is from this adaptability that the model's scalability arises. To fully exploit this adaptable topology, in this work a method is proposed which calculates tuning network topologies composed of the minimum number of analysis modules required to provide effective dynamic tuning. The proposed model has been implemented in the form of ELASTIC, an environment for large-scale dynamic tuning. ELASTIC presents a plugin architecture, which allows different performance analysis and tuning strategies to be applied. Using ELASTIC, experimental evaluation has been carried out on a synthetic and a real parallel application. The results show that the proposed model, embodied in ELASTIC, is able to not only scale to meet the demands of dynamic tuning over thousands of processes, but is also able to effectively improve the performance of these applications.

Imprimir

Modificar Consulta Nueva Consulta

Versión 4.4