公开(公告)号：JPH09330304A

公开(公告)日：1997-12-22

申请号：JP14244096

申请日：1996-06-05

Applicant: IBM

Inventor： KOMATSU HIDEAKI ,  OGASAWARA TAKESHI

IPC: G06F13/00 ,  G06F15/173 ,  G06F15/163

Abstract: PROBLEM TO BE SOLVED: To generate an optimum communication schedule when data are transmitted and received between processors which constitute parallel computers or a decentralized multiprocessor system. SOLUTION: The processors which communicate with one another are classified by groups (step 40), nodes are provided for the groups, one to one, and a communication graph having respective sides made to correspond to the communication is generated (steps 4 and 3). Graph are generated by the distances between the nodes. Then the communication graphs by the distances between the nodes are made to correspond to communication steps of the inter- processor communication (step 46). The communication to be made is totally grasped with the communication graphs and the sides of the communication graphs are made to correspond to the inter-processor communication which is made in a certain communication step, so the communication can be optimized.

公开(公告)号：JPH0830561A

公开(公告)日：1996-02-02

申请号：JP15494294

申请日：1994-07-06

Applicant: IBM

Inventor： KOMATSU HIDEAKI ,  ISHIZAKI KAZUAKI ,  GODA OSAMU

IPC: G06F15/16 ,  G06F9/45

Abstract: PURPOSE: To attain the high speed of the execution of a program by introducing communication dependence expressing data dependence between processors. CONSTITUTION: A data dependence analyzing part 304 operates normal data dependence analysis and generates a data dependence vector for a loop decided by an object loop setting part 302. Then, an iteration space dividing part 306 divides all the areas (iteration spaces) of an index executed by the loop. Then, a communication analyzing part 308 analyzes which data should be read for executing the iteration spaces divided for being executed in parallel by each processor, compares it with it own, and decides the area of the data for communication. Moreover, a communication dependence analyzing part 310 processes the calculation of a communication vector, the calculation of a communication dependence vector, the judgement of the communication configuration of an operand, and the judgement of the execution configuration of the loop in this procedure.

公开(公告)号：JPH07152576A

公开(公告)日：1995-06-16

申请号：JP29626993

申请日：1993-11-26

Applicant: IBM JAPAN

Inventor： KOMATSU HIDEAKI

IPC: G06F9/45 ,  G06F17/16

Abstract: PURPOSE:To provide the widely applicable in-line expansion method without preparing individual templates or macros for various composite formulas of array functions and array operation in the programming language having array functions such as APL and FORTRAN 90. CONSTITUTION:A TIF (built-in array transformation function) is subjected to in-line expansion as follows; the TIP is made into a virtual array and the TIF is subjected to macro expansion at this time to generate mapping to an actual array (1), and the virtual array is subjected to in-line expansion as it is (2), and the virtual array is converted back to the actual array (3). In this method, the TIF formula is analyzed at the time of compilation to generate a template for in-line expansion. This temperate for in-line expansion is mapping to the actual array, and macro processing is used to generate it. By this method, arbitrarily combined TIFs are subjected to in-line expansion by only preparing one macro template for one TIF.