Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 58120 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Sarbazi Azad, Hamid
- Abstract:
- In recent years, graphical processing units (GPUs) have gained increasing popularity for scientific calculations, performing solutions, training and developing machine learning algorithms and many other applications. These processors, designed to run a large number of threads simultaneously, have the potential to execute mentioned applications faster than central processing unit (CPUs). Many of the modern GPU programs perform numerous memory accesses, GPUs are included with caches. Although for many applications, caches alone can bring limited improvements, in recent studies, use of prefetching for improving benefits of GPU caches has been investigated and has shown promising results. These studies usually add a prefetching unit to GPU hardware. This approach requires new processors to be designed and would take up some of the resources of the chips. In this study, we present a software based approach that enables developers to effortlessly add prefetching to their GPU applications. This approach that is usable on all modern GPUs leads to significant improvement to cache hit rate and considerably reduces program stalls. We have observed improvements of up to 80 percent for cache hit rate and stall reductions of up to 65 percent, with the average hit rate improvement of 27 percent and average stall reduction of 7 percent
- Keywords:
- Graphics Procssing Unit (GPU) ; Prefetching ; Cache Memory ; Hierarchical Memory Structure ; Increasing Efficiency
Digital Object List
محتواي کتاب
Bookmark