32
Single Program, Multiple Data
A program multiple data (SPMD) in the science cloud is one of the techniques is to achieve computing parallel it used to be. This method, set method, several ordered multiple data (MIMD). In this method, split and in parallel on several processors with the input of parallel to achieve a faster result Run can be. Techniques SPMD the most commonly used method in parallel processing. This method is also the prerequisite concepts of research, such as messages, active memory and distributed shared.
techniques SPMD usually to method transfer messages the program in a computer architecture using memory distributed are applied, a computer architecture distributed from a number of independent computer composed that each one of them, node (Node) call. Each node program begins and through send and receive messages with other nodes to communicate. With the use of calling, send/receive. The possibility of implementation commands, such as the impossibility of sync (Barrier Synchronization), and also with messages there. This message, with methods of data transfer such as TCP/IP on Ethernet or proprietary methods very high speed such as Myrinet and domestic connections supercomputer implementation of. Section Serial program through the calculation of pointers is quite similar in all the nodes is implemented, and virtually of the methods such as calculating the result in one node and send the result to the other nodes don't use.
in a machine with memory shared etc. can be applied to a message, re-submit its contents in an address space shared access. This method is usually the most optimal procedures in the planning computers with address space shared and the number of processors is high. Especially in the car, NUMA is that memory belongs to any processor has access to memory of other processor needs more time. Techniques SPMD machines, the memory shared, usually by the process of Standard very heavy implement.
unlike SPMD, etc. with memory shared to which multi-processor parallel (SMP) is also applicable. to the programmer the possibility of using a memory space shared gives, in which process, in parallel on different processors, and separate routes run can be. Program in a processor starts and its implementation to the regions split can be that at the time of the start of the parallel instructions are implemented. In the areas of parallel and ... processor, a program with a different data run. An example. Loop Do .... Loop. in which different processors are busy implementing the different parts of the array within the loop are responsible for. At the end of the loop, the results of the sync are at this time only one processor is busy and other processors are in waiting mode. Interface the current standard for with shared memory OpenMP is called, and usually using a process of non- to the name of the yarn (threads) implementation.
computers, the current possibility of combining many rules and modes parallel to achieve the maximum amount of productivity to the programmer. A program distributed memory that MPI uses can be on a very large number of nodes run. each node can be a computer with memory shared and to be parallel in several processors by the OpenMP run. Within each processor, etc. recipes vector SIMD (usually automatically by the compiler are run) and instructions Superscalar (usually by the CPU for an independent run can be) such as Pipelining and the use of several units operating separate parallel to achieve maximum speed in each processor are used.
techniques SPMD first in 1983 by Michel Auguin and Francois Larbey in computer parallel OPSILA used and then in 1984 by Bride Darema at IBM for the machine, above parallel, such as RP3 was used. Based on one of the documentation, non-published IBM in the 1980's from non-standard lot for the transfer of the messages is remembered, but first, the standard SPMD, PVM has been. The dominant standard today, MPI.
the technique of parallelization of Cray the previous generation, OpenMP can be....
