EXPERIENCE IN EFFICIENT OPTIMIZATION AND PORTING OF NUMERICAL APPLICATIONS TO UP-TO-DATE PARALLEL COMPUTERS

VI.V. Voevodin
VANT. Ser.: Mat. Mod. Fiz. Proc 1997. Вып.1. С. 65.

      It is commonly adopted opinion that parallel computers are very hard to be used efficiently. The announced values of performance are high but-in practice obtained results look less attractive.-Many scientists have already experienced serious difficulties in programming for a such kind of computers.. These, difficulties are easy to understand since a lot of new and unusual problems occur almost everywhere: what is parallelism, how to; detect it; how one. should express parallelism in a high level language, how.to make a performance closer to a peak (or theoretical) performance of a computer, can we use efficiently a large existing resource of algorithms and programs, and how to guarantee that the efforts which have been spent to develop parallel software will not have to be applied again and again with the emergence of computers.
      This report presented different aspects of program analysis and optimization for up-to-date parallel computers. The first part describes the theoretical basis of so called V-Ray technology directed to comprehensive analysis of programs. This technology is a set of mathematical methods and algorithmic approaches designed to facilitate investigation and transformation of the structure of serial algorithms and programs. The V-Ray technology is developed on the basis of the strict theory and provides a basis for resolving the whole scope of problems related to mapping of real world applications to parallel computers starting from the visual analysis of program structure and detection of data dependencies, description of the total resource of parallelism, search for potential bottlenecks in programs, up to analysis of possible data distributions and analysis of data locality.
      The second part outlines the functionality of the V-Ray system designed on the principles of the V-Ray technology. The main goal of the system is to provide powerful tools for performing comprehensive analysis of program structure. Since the reasons for low performance are different, the system supports investigation and evaluation of different properties of programs on all levels: from an interprocedure level up to individual loop iterations.
      The final part of the report describes our experience in porting several programs to massively parallel (CRAY T3D, IBM SP2) and vector-parallel (CRAY Y-MP C90) computers. Results and the main stages of this process are shown. For instance, , the detection of the total resource of parallelism hidden in the program for modeling large-scale magnetic fields in galaxies enabled us to attain 7-fold speedup on the CRAY Y-MP C90 computer.