From: Luc.Vereecken@chem.kuleuven.ac.be (Luc Vereecken) Newsgroups: comp.parallel.mpi,comp.parallel.pvm Subject: Re: from PVM to MPI Date: Sat, 05 Dec 1998 16:35:24 GMT Organization: KULeuvenNet Message-Id: <36695fb5.84081387@news.kuleuven.ac.be> References: <366804F6.84@pvmcube4.snu.ac.kr> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: 8bit Cache-Post-Path: marvin!unknown@pcjp1.chem.kuleuven.ac.be Xref: ukc comp.parallel.mpi:4404 comp.parallel.pvm:7836 On Sat, 05 Dec 1998 00:51:18 +0900, "Lee, Bo-sung" wrote: > >> The communication structure of the application is >> rather simple master-slave. >> All the slaves are started once, at the beginning of computation. >Most MPI don't allow master-slave model. So, you need to convert >master-slave apps to SPMD model. It's simple I think. I know of no MPI implementation that doesn't allow a master-slave model. Master-slave does not imply that one program must be able to start other, only that one process behaves differently than the other, acting as a master. This is easily done by either choosing e.g. process zero as a master, or write the program as a MPMD program (as I am doing for a project at the moment). The main problem with porting PVM to MPI is if your PVM application alters the type/number of processes during the run. With MPI, all processes must be started at the start of the job. (But that will change with MPI-2, when it is implemented). Luc Vereecken