Loading...
Please wait, while we are loading the content...
Similar Documents
The refined-language approach to compiling for parallel supercomputers
| Content Provider | Semantic Scholar |
|---|---|
| Author | Dietz, Henry G. |
| Copyright Year | 1987 |
| Abstract | To produce efficient code for a program which is to be executed using a particular parallel supercomputer, it is vital that the compiler employ detailed knowledge of both the algorithm and the target machine. However, the specifications of algorithms in conventional, sequential, languages often are not sufficiently precise; information which the programmer knew, and which the compiler must use to generate efficient parallel code, is either hidden or lost because the language does not provide a way to express that information. In addition, although many compilation technologies have been developed to support specific classes of speedup-oriented parallel supercomputer, relatively little has been done to unite these different compilation technologies in a consistent view of code generation for all parallel machines; this has often led to extremely clever code generation resulting in code which is inappropriate for the target machine and therefore slow. The refined-language approach is an holistic strategy for solving the problem of compiling, and programming, for speedup-oriented parallel supercomputers. So that information relevant to parallelization of an algorithm is not lost when the program is written, we provide guidelines for "refining" the definition of programming languages to permit the programmer to explicitly state, as part of his/her programs, this information. We further propose a framework for automatic parallelization and related compiler technologies which spans many classes of architecture and highlights the close relationship between conventional optimization and optimizing parallelization. Several new parallelization techniques are embedded in this discussion. A prototype refined-C (RC) compiler, based on full ANSI C and targeted to a MIMD supercomputer, has been built. Although this compiler has not been thoroughly benchmarked, it has demonstrated both that language refinement greatly improves the speed and accuracy of compile-time analysis and that, at least in some cases, efficient parallel code can be generated for constructs which would have been sequential using other techniques. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://aggregate.ee.engr.uky.edu/REFINED/thesis.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
