|
NAMEParallel::Pvm - Perl extension for the Parallel Virtual Machine (PVM) Message Passing SystemSYNOPSISuse Parallel::Pvm; DESCRIPTIONThe PVM message passing system enables a programmer to configure a group of (possibly heterogenous) computers connected by a network into a parallel virtual machine. The system was developed by the University of Tennessee, Oak Ridge National Laboratory and Emory University.Using PVM, applications can be developed which spawns parallel processes onto nodes in the virtual machine to perform specific tasks. These parallel tasks can also periodically exchange information using a set of message passing functions developed for the system. PVM applications have mostly been developed in the scientific and engineering fields. However applications for real-time and client/server systems can also be developed. PVM simply provides a convenient way for managing parallel tasks and communications without need for rexec or socket level programming. As a utility, PVM enables an organisation to leverage on the computers already available for parallel processing. Parallel applications can be started during non-peak hours to utilise idle CPU cycles. Or dedicated workstation clusters connected via a high performance network like ATM can be used for high performance computing. It is recommended that you read the PVM manual pages and the book "PVM: Parallel Virtual Machine, A users's guide and tutorial for networked parallel computing". Both the PVM system and the book can be obtained from the HTTP address http://www.epm.ornl.gov/pvm. For the rest of this document we will provide a tutorial introduction to developing PVM applications using perl. The interface for some of the PVM functions have been changed of course to give it a more perl-like feel. Remember think perl think parallel! Good Luck! Environment VariablesAfter installing PVM on your computer, there are two mandatory environment variables that have to be set in your .login or .cshrc files; PVM_ROOT and PVM_ARCH. PVM_ROOT points to the base of the PVM installation directory, and PVM_ARCH specifies the architecture of the computer on which PVM is running. An example of how this can be set for csh is shown below,setenv PVM_ROOT /usr/local/pvm3 setenv PVM_ARCH `$PVM_ROOT/lib/pvmgetarch` Setting up your rsh permissionIn order for PVM applications to run, rsh permission has to be enabled. This involves creating a .rhosts file in your HOME directory containing, for each line, the host and account name you wish to allow remote execution privillages. An example .rhosts file to allow a PVM application to remotely execute on the host onyx and prata using the account edward is shown below,onyx edward prata edward Configuring your parallel virtual machineParallel process management and communications is handled by a set of distributed deamons running on each of the nodes of the virtual machine. The daemon executable, pvmd, is started when a computer is added to the virtual machine. A computer can be added to the virtual machine either statically in a console program or using a hostfile, or dynamically within the application code itself.The first method of configuring your virtual machine is to use the console program $PVM_ROOT/lib/pvm. Run it from the command prompt. The console program will first add the local host into the virtual machine and display the prompt pvm> To add a host, eg onyx, as a node in your parallel virtual machine, simply type pvm> add onyx To display the current virtual machine configuration type pvm> conf which will display node information pertaining to the host name, host id, host architecture, relative speed and data format. The console program has a number of other commands which can be viewed by typing help. The second method of configuring your virtual machine is to use a hostfile. The hostfile is simply an ASCII text file specifing the host names of the computers to be added into your virtual machine. Additional options may be also be defined for the nodes pertaining to the working directory, execution path, login name, alternative hostname etc. A simple example of a hostfile is shown below. * wd=$HOME/work ep=$HOME/bin onyx prata.nsrc.nus.sg laksa ep=$HOME/perl5/bin In the above example hostfile we are adding the hosts onyx, prata.nsrc.nus.sg and laksa into the virtual machine. We are also specifying the working directory, wd, in which we want our application to run, and the execution path, ep, in which we want PVM to look for executables. The * in the first line defines a global option for all the hosts specified after it. We can however provide an option locally to over-ride this global option. This is seen for the host laksa where we have specified its execution path to be $HOME/perl5/bin instead of the $HOME/bin. The third method of configuring your virtual machine is to call the functions Parallel::Pvm::addhosts or Parallel::Pvm::delhosts within your application. You must still start your master pvmd daemon first. This can be achieved by starting pvm and typing quit or simply typing echo quit | pvm The PVM application can then be started where we can add the hosts prata and laksa by calling Parallel::Pvm::addhosts("prata","laksa"); Or we can delete a host from our configuration by calling Parallel::Pvm::delhosts("laksa"); PVM also provides a function, Parallel::Pvm::conf, to query the configuration of the parallel virtual machine. An example code to check the current configuration is shown below. ($info,@conf) = Parallel::Pvm::conf ; if ( $info == PvmOk ){ foreach $node (@conf){ print "host id = $node->{'hi_tid'}\n"; print "host name = $node->{'hi_name'}\n"; print "host architecture = $node->{'hi_arch'}\n"; print "host speed = $node->{'hi_speed'}\n"; } } Enrolling a task into PVMA task has to expilictly enroll into PVM in order for it to be known by other PVM tasks. This can often be done by the call$mytid = Parallel::Pvm::mytid ; where $mytid is the task id, TID, assigned by the PVM system to the calling process. Note however that calling any PVM function in a program will also enroll it into the system. Spawning parallel tasksA PVM application can spawn parallel tasks in your parallel virtual machine. Assuming there is exists an executable called client, we can spawn four client tasks in our virtual machine by calling($ntask,@tids) = Parallel::Pvm::spawn("client",4); For each of the four spawned processes, the PVM system first allocates a host node and looks for the executable in the execuation path of that host. If the executable is found it is started. The task which called the Parallel::Pvm::spawn is known as the parent task. The number of children tasks which are actually spawned by Parallel::Pvm::spawn is returned in the scalar $ntask. The @tids array returns the task id, TID, of the spawned children tasks which will be useful later for communicating with them. A TID < 0 indicates a task failure to spawn and can be used to determine the nature of the problem. Eg. foreach $tid (@tids){ if ( $tid < 0 ){ if ( $tid == PvmNoMem ) warn "no memory ! \n"; }else if ( $tid == PvmSysErr ){ warn "pvmd not responding ! \n"; } ... } } For more sophisticated users, Parallel::Pvm::spawn may be given additional argument parameters to control how/where you want a task to be spawned. For example, you can specifically spawn client in the internet host onyx.nsrc.nus.sg by calling Parallel::Pvm::spawn("client",1,PvmTaskHost,"onyx.nsrc.nus.sg"); Or you can spawn client on host nodes only of a particular architecture, say RS6K workstations, by calling Parallel::Pvm::spawn("client",4,PvmTaskArch,"RS6K"); Also, if the spawned remote executable requires an argument argv, you can supply this by calling Parallel::Pvm::spawn("client",4,PvmTaskArch,"RS6K",argv); Note that tasks which have been spawned by using Parallel::Pvm::spawn do not need to be explicitly enrolled into the pvm system. Exchanging messages between tasksMessages can be sent to a task enrolled into PVM by specifying the example code sequenceParallel::Pvm::initsend ; Parallel::Pvm::pack(2.345,"hello dude"); Parallel::Pvm::pack(1234); Parallel::Pvm::send($dtid,999); In our example we first call Parallel::Pvm::initsend to initialize the internal PVM send buffer. We then call Parallel::Pvm::buffer to fill this buffer with a double (2.345), a string ("hello dude"), and an integer (1234) <b>Actually, currently all arguments are converted to strings</b>. Having filled the send buffer with the data that is to be sent, we call Parallel::Pvm::send to do the actual send to the task identifed by the TID $dtid. We also label the sending message to disambiguate it with other messages with a tag. This is done with the 999 argument in Parallel::Pvm::send function. For the destination task, we can receive the message sent by performing a blocking receive with the function Parallel::Pvm::recv. A code sequence for the above example on the recipent end will be if ( Parallel::Pvm::recv >= 0 ){ $int_t = Parallel::Pvm::unpack ; ($double_t,$str_t) = Parallel::Pvm::unpack ; } Note that we must unpack the message in the reverse order in which we packed our message. In our example Parallel::Pvm::recv will receive any message sent to it. In order to selectively receive a message, we could specify the TID of the source task and the message tag. For example, $tag = 999; Parallel::Pvm::recv($stid,$tag) ; Caveats: Messages may not contain the vertical tab character "\v". If you pass messages to programs written in other languages, you need to know that "Parallel::Pvm::pack" packs everything as strings (with "pvm_packstr"). Other message passing functions that you may find useful are Parallel::Pvm::psend, Parallel::Pvm::trecv, Parallel::Pvm::nrecv and Parallel::Pvm::precv. Parallel I/ONote that the file descriptors in a parent task are not inherented in the spawned children tasks unlike fork. By default any file I/O will be performed in the working directory specified in the hostfile if no absolute path was provided for the opened file. If no working directory is specified, the default is the $HOME directory. For directories which are not NFS mounted, this would mean that each task performs its own separate I/O.In the case of tty output, tasks which are not started from the command prompt will have their stdout and stderr directed to the file pvml.<uid>. This may be redirected to a parent task by calling Parallel::Pvm::catchout; for stdout or Parallel::Pvm::catchout(stderr); for stderr. You can direct the stdout or stderr output of a task to another TID , other then its parent, by calling Parallel::Pvm::setopt(PvmOutTid,$tid); Incorporating fault toleranceThe function Parallel::Pvm::notify can be used to incorporate some fault tolerance into your PVM application. You may use it to ask the PVM to monitor the liveliness of a set of hosts or tasks during the execution of a PVM application. For example you can instrument your application to monitor 3 tasks with TID $task1, $task2, and $task3, by using the code segments@monitor = ($task1,$task2,$task3); Parallel::Pvm::notify(PvmTaskExit,999,@monitor_task); ... if ( Parallel::Pvm::probe(-1,999) ){ $task = Parallel::Pvm::recv_notify ; print "Oops! task $task has failed ... \n" ; } If either $task1, $task2 or $task3 fails, the notification will take the form of a single message with the tag 999. The message content will inform you of the TID of the failed task. A similar scheme may be employed for the notification of host failures in your parallel virtual machine. Client/Server exampleClient:use Pvm; use File::Basename; ... # Look for server tid and assume # server name is 'service_provider' @task_list = Parallel::Pvm::tasks ; foreach $task (@task_list){ $a_out = $task->{'ti_a_out'} ; $base = basename $a_out ; if ( $base eq 'service_provider' ) $serv_tid = $task->{'ti_tid'} ; } # This is just one way (not necessarily the # best) of getting a server tid. # You could do the same thing by reading # the server tid posted in a file. ... # send request for service Parallel::Pvm::send($serv_tid,$REQUEST); # receive service from server Parallel::Pvm::recv(-1,$RESPONSE); @service_packet = Parallel::Pvm::unpack ; ... Server: while(1){ ... if ( Parallel::Pvm::probe(-1,$REQUEST) ){ # a service request has arrived ! $bufid = Parallel::Pvm::recv ; ($info,$bytes,$tag,$stid) = Parallel::Pvm::bufinfo($bufid) ; if ( fork == 0 ){ # fork child process to handle service ... # provide service Parallel::Pvm::initsend ; Parallel::Pvm::pack(@service); Parallel::Pvm::send($stid,$RESPONSE); # exit child process exit ; } } ... } PVM groupsThe PVM dynamic group functions have not completely been ported to Perl yet. We do not support pvm_scatter, pvm_gather, and pvm_reduce currently. This is connected to the limited datatype support in the rest of the Perl interface.The group functions provide facilities for collecting processes under a single group label, and applying aggregate operations onto them. Examples of these functions are Parallel::Pvm::barrier, Parallel::Pvm::reduce, Parallel::Pvm::bcast etc. One of our concerns is that these group functions may be changed or augmented in the future releases of PVM 3.4*. FUNCTIONS
AUTHORSEdward Walker, edward@nsrc.nus.sg, National Supercomputing Research Centre, SingaporeDenis Leconte, denis_leconte@geocities.com Ulrich Pfeifer, pfeifer@wait.de SEE ALSOperl(1), pvm_intro(1PVM)
Visit the GSP FreeBSD Man Page Interface. |