Loading...
Please wait, while we are loading the content...
Um middleware independente da plataforma para computação paralela
| Content Provider | Semantic Scholar |
|---|---|
| Author | Mourão, Diogo |
| Copyright Year | 2011 |
| Abstract | The widespread adoption of multi-core processors requires programming models that allow to express parallelism in a simple way, without exposing low level details about concurrency management. Nevertheless, despite multi-core processors becoming the de-facto standard since 2006, commonly used programming languages and their respective compilers and runtime systems remain the same. This is reflected not only at the applications performance level but also at their development productivity. As such, it is then desirable to create new solutions that offer high level abstractions allowing to express parallelism in a simple way, and to separate the application logic from the concurrency management at the same time. In this context, this thesis proposes a independent middleware platform to support parallel applications execution by providing common functionality such as: task and data parallelism, communication and concurrency control. It aims to be sufficiently generic to be able to support the development of a great variety of concurrent and parallel applications, as well as supporting programming languages runtime systems. Its architecture is inspired in that of operating systems, in the sense that, not only having a well defined interface to the programmer, based on places, it also defines an interface, based on driver concept, to support multiple implementations of the needed functionalities to execute parallel applications. In an effort to ease application development, an annotation mechanism is available which allows the programmer to express parallelism in his applications by just annotating the source code. Our middleware already implements support for shared memory architectures but its design allows future support for distributed memory architectures, to which work is already being done. The present implementation is evaluated performance wise through the NAS Parallel Benchmarks [NAS] and Java Grande Benchmark Suite [EPC] benchmarks. The results confirm that using our middleware incurs in some overhead, as expected. However, it is compensated by the abstractions offered which greatly simplify |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://run.unl.pt/bitstream/10362/6322/1/Mourao_2011.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
