|
Identifier |
000360801 |
Title |
Locality management in task-based parallel programming models |
Alternative Title |
Διαχείρηση τοπικότητας σε βασισμένα σε εργασίες παράλληλα προγραμματιστικά μοντέλα |
Author
|
Κούκος, Κωνσταντίνος Ιωάννη
|
Thesis advisor
|
Νικολόπουλος, Δημήτριος
Μπίλας, Άγγελος
|
Abstract |
In this thesis we discuss the oerformance impact of cache locality on multicore x86 architectures. Our goal is to improve performance and efficiency on these systems. We suggest a novel task-based, parallel runtime system (TPR), that allows programmer to create custom scheduling policies and perform guided prefetching on the task arguments with minimal training effort requirements. TPR is targeted on very low overhead task creation, issue and synchronization. This system is designed to facilitate the programmer with utilities such as explicit data annotations to improve locality via prefetching and task binding to implement custom scheduling policies. Using that system we explore the parameters that affects locality in a major degree and the performance impact of each one. We coclude that the most important locality parameters are memory layout, task size, scheduling policy and SW prefetching. For the evaluation we use both linear algebra kernels and real applications. The kernels discussed are GEMM, LU, Cholesky and Jacobi. The applications are PBPI, FixedGrid, and FFT. Finally we compare our work with other based runtime systems and hand tuned applications written using pthreads or MPI. We manage an average performance improvement over other runtime systems from 1% to 40% depending on the application. The contribution of this work is to set a minimal subset of requirements for a runtime system to efficiently handle locality and to categorize applications and assign their primary optimization technique.
|
Language |
English |
Subject |
Locality |
|
Performance improvement |
|
Runtime systems |
|
Βελτιστοποίηση απόδοσης |
|
Συστήματα χρόνου εκτέλεσης |
|
Τοπικότητα |
Issue date |
2010-11-19 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Views |
577 |