PARALLEL PROGRAMMING IN PARALLEL VIRTUAL MACHINE Fayon Atkinson

PARALLEL PROGRAMMING IN PARALLEL VIRTUAL MACHINE Fayon Atkinson 1

History � Oak Ridge National Laboratory in 1989 � University of Tennessee in 1991 � Said to have also been revised at Emory University later. 2

PVM Explained � � � Applications on a set of heterogeneous computers connected by a network. Appears logically as single parallel computer to its users. Distributed computing architecture is created from the software’s from a parallel connected system. Makes parallel programming straight forward and efficient Accurate computation Use to teach science medicine & engineering 3

Environment � Two component: a library of PVM routines, and a daemon that should reside on all the hosts in the virtual machine. � Hosts: computer nodes which could be uniprocessors, multiprocessors or clusters running the software. 4

Structure � Star graph structure �Supervisor –workers model � Tree structure �Hierarchy model 5

Supervisor–Workers Model Supervisor is the middle node in the star � The supervisor is the initiating task � One supervisor � � interact with the user, activating the workers, assigning work to the workers, and collecting results from the workers. � Many workers � Calculations � independent or dependent � dependent workers communicate with each other before sending supervisors the computations 6

Hierarchy Structure Root of the tree is the top supervisor � Workers can create a set of new workers on another level � Parent is the task that creates another task � 7

Task Creation � task can be started manually or can be spawned from another task � pvm_spawn() All tasks are given a unique identifier when created (TID) � To create a child: � � Child machine is started � path to the executable file on the specified machine. � number of copies of the child to be created. � array of arguments to the child task(s). 8

Task Creation � Functions: � mytid = pvm_mytid(); - running task can retrieve its own TID � pvm_spawn(…, …, &tid) - array containing the TIDs of the children created by this call will be returned � my_parent_tid = pvm_parent(); - returns the parent. Will return Pvm. No. Parent for initial tasks � daemon_tid = pvm_tidtohost(id); - determining on which host a given task is running � num = pvm_spawn(Child, Arguments, Flag, Where, How. Many, &Tids). 9

Task Groups Grouping tasks is common Useful for when a collective operation is performed on only a subset of the tasks, such as sending a message to only members of a group. � Functions: � � � i = pvm_joingroup(group_name) – joining a group called group_name � info = pvm_lvgroup(group_name) – leaving a group called group_name � pvm_gsize() – retrieve size of group � pvm_gettid() – retrieve TID if a task given instance number and group name, � pvm_getinst() – retrieve instance number of a given task and group name. 10

Task Communication � Sending Steps: 1. 2. 3. � A send buffer must be initialized. The message is packed into this buffer. The completed message is sent to its destination(s). Receiving Steps: 1. 2. The message is received. The received items are unpacked from the receive buffer. 11

Thank You! 12