From: Bum-Seok Hyun Newsgroups: comp.parallel.mpi Subject: Re: MPI_REDUCE Date: Tue, 12 Jan 1999 10:08:22 +0900 Organization: Seoul National University, Republic of Korea Message-Id: <369AA085.870D6A39@gong.snu.ac.kr> References: <3699D811.B75AE147@nada.kth.se> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Bjorn Sjogreen wrote: > I can not find anywhere in the standard document, whether a call to > MPI_Allreduce (or Reduce) with > the same variable as both input and output buffer will give the > expected result. For example > what does the standard say about the following call > MPI_Allreduce( buf, buf, n, MPI_DOUBLE, MPI_MAX, comm ) > regards > Bjorn Sjogreen > KTH, Stockholm, Sweden Hi! In the book "Parallel programming with MPI" , the standard will not allow to use same variable in MPI_REDUCE, it says we should use something like, MPI_REDUCE(local variable, global variabl, ...) But something's weird here when I'm using REDUCE. In LAM MPI on intel LINUX, and on Cray T3E, there were no problems using, MPI_REDUCE(error, error,...) In MPICH on Intel or Alpha Linux, it showed error message... So I had to change to MPI_REDUCE(error, errorg,...) One more strange thing in MPICH on Alpha-Linux, When I had to use more than one variable to REDUCE, like, MPI_REDUCE(error, errorg,...) MPI_REDUCE(force, forceg,...) MPICH produced device error after starting to run code... (But, LAM MPI didn't complain anything, though) Regards. -- ------------------------------ Bum-Seok Hyun Ph.D. Candidate Aerodynamic Simulation Lab. Dept.of Aerospace Engineering Seoul National University Office : +82-2-880-1903 Fax : +82-2-887-2662 mailto:dominic@gong.snu.ac.kr