User:Szha/Notes/CMSC714/note0906

1. notes
2. first assignment MPI next week
3. accounts emailed next week for UMIACS bug cluster and for SMP
4. check readings page to see question sending assignment:w
5. 2-4 questions
6. 6PM day before lecture

PVM

run on everything

parallel hardware: SMP, MPPs, Vector machines

network of workstations: ATM, ethernet,

• UNIX machines and PCs running win32 API

works on a heterogenous collection of machines

• handles type conversion as needed

provides two things

message pasing library

• point-to-point messages
• synchronization: barriers, reductions

OS support

• process creation (pvm_spawn)

PVM environment (UNIX)

one PVMD (daemon) per machine

all processes communicate through pvmd (by default)

PVM message passing

all messages have tags

an integer to identify the message

defined by the user

messages are constructed, then sent

pvm_pk{int,char,float}(*var, count, stride)

pvm_unpk{int,char,float} to unpack

all processes are named based on task ids (tids)

local/remote processes are the same

primary message passing functions

pvm_send(tid,tag)

pvm_recv(tid,tag) # tid to receive from, could be anywhere

PVM process control

creating a process

pvm_spawn(task,argv,flag,where,ntask,tids)

task is name of program to start

flag and where provide control of where tasks are started

ntask determines how many copies are started

program must be installed on each target machine

returns number of tasks actually started

ending a task

pvm_exit

does not exit the process, just the PVM machine

info functions

pvm_mytid() get the process task id

PVM group operations

group is the unit of communication

a collection of one or more processes

processes join group with pvm_joingroup("<group name>")

each process in the group has a unique id

• pvm_gettid("<group name>")

barrier

can involve a subset of the processes in the group

pvm_barrier("<group name>"), count)

reduction operations

pvm_reduce( void (*func)(), void *data, int count, int datatype, int msgtag, char *group, int rootinst)

• result is returned to rootinst node
• does not block (does not wait until root has the return value, just indicate valid buffer value)

pre-defined funcs: PvmMin, PvmMax, PvmSum, PvmProduct

PVM performance issues

messages have to go through PVMD

can use direct route option to prevent this problem

packing messages

semantics imply a copy (at least)

extra function call to pack messages

heterogenous support

information is sent in machine independent format

has a short circuit option for known homogenous comm.

• passes data in native format then

sample PVM program

fragment from main

myGroupNum = pvm_joingroup("ping-pong");
mytid = pvm_mytid();
if (myGroupNum == 0) {
    pvm_catchout(stdout);
    okSpawn = pvm_spawn(MYNAME, argv, 0, "", 1, &friendTid);
    if (okSpawn != 1) {
        printf("can't spawn");
        pvm_exit();
        exit(1);
    }
} else {
    friendTid = pvm_parent();
    tids[0] = friendTid;
    tids[1] = mytid;
}
pvm_barrier("ping-pong", 2);

if (myGroupNum == 0) {
    for (i = 0; i < MESSAGESIZE; i++) {
        messsage[i] = '1'l
    }
}

for (i = 0; i<ITERATIONS; i++) {
    if( myGroupNum == 0) {
        pvm_initsend(PvmDataDefault);
        pvm_pkint(message,MESSAGESIZE, 1); // stride 1 indicates no gap
        pvm_send(friendTid, msgid);
        pvm_recv(friendTid, msgid);
        pvm_upkint(message,MESSAGESIZE,1);
    }
    else {
        pvm_recv(friendTid,msgid);
        pvm_upkint(message,MESSAGESIZE, 1);
        pvm_initsend(PvmDataDefault);
        pvm_pkint(message, MESSAGESIZE, 1);
        pvm_send(friendTid,msgid);
    }
}
pvm_exit();
exit(0);