CUBIX: Programming hypercubes without programming hosts

Conference · Wed Jan 01 04:00:00 EST 1986

OSTI ID:6905604

Salmon, J

Typically, application programs for hypercubes consist of two parts, a master process running on the host and a server running in the nodes of the hypercube. CUBIX adopts a different viewpoint. Once a program is loaded into the hypercube, that program assumes control of the machine. The host process only serves requests for operating system services. Since it is no more than a file server, the host program is universal; it is unchanged from one application to the next. This programming model has some important advantages; (1) program development is easier because it is not necessary to write a separate program for the host, (2) hypercube programs are easier to develop and debug because they can use standard I/O routines, rather than machine dependent system calls, (3) hypercube programs can often be run on sequential machines with minimal modification, (4) versions of CUBIX exist for both crystalline and amorphous applications. In crystalline applications operating system requests occur synchronously. Requests are either ''singular'' or ''multiple'', according to whether all nodes request the same or distinct actions. In amorphous applications requests occur completely asynchronously. The host process serves each request independently. Care is taken so that many processors may simultaneously access the same file.

• Research Organization: California Inst. of Tech., Pasadena (USA). Dept. of Physics
• DOE Contract Number: FG03-85ER25009
• OSTI ID: 6905604
• Report Number(s): DOE/ER/25009-378; CONF-8609173-6; ON: DE87005030
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990210* -- Supercomputers-- (1987-1989)
C CODES
COMPUTER CODES
COMPUTERS
DIGITAL COMPUTERS
PARALLEL PROCESSING
PROGRAMMING
SUPERCOMPUTERS