ETD

Archivio digitale delle tesi discusse presso l'Università di Pisa

Tesi etd-11152015-211525


Tipo di tesi
Tesi di laurea magistrale
Autore
PARISI, CLAUDIO
URN
etd-11152015-211525
Titolo
Optimization of Parallel Computations on Heterogeneous GPU-Based Systems
Dipartimento
INFORMATICA
Corso di studi
INFORMATICA E NETWORKING
Relatori
relatore Prof. Danelutto, Marco
correlatore Prof. Kessler, Christoph
Parole chiave
  • GPGPU
Data inizio appello
04/12/2015
Consultabilità
Completa
Riassunto
In this master thesis, we design and implement MultiStream: a solution that extends the existing data parallel skeleton library SkePU with NVIDIA CUDA Streams to overlap main memory – device memory data transfers with CUDA Kernel executions.
We show the benefits of this approach using a task-parallel framework, FastFlow, on-top of SkePU.
Finally, we compare the MultiStream extended SkePU to an ad-hoc solution to discuss the tradeoffs between the level of abstraction and the maximum achievable performance.
File