From: Xingfu Wu Newsgroups: comp.lang.fortran,comp.parallel.pvm,comp.parallel.mpi Subject: Re: superlinear speedup, (was: fortran77 programming) Date: Wed, 16 Dec 1998 12:55:27 -0600 Organization: Department of Computer Science, LSU Message-Id: <3678021F.41B91CBF@bit.csc.lsu.edu> References: <3660D892.E1EDBF12@est.it> <3676AC48.18DE0329@bit.csc.lsu.edu> <756qn6$304$1@news.uky.edu> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Xref: ukc comp.lang.fortran:62238 comp.parallel.pvm:7883 comp.parallel.mpi:4447 I agree that "superlinear convergence" is actually misleading. I thought that it is not useful to only pursue the superlinear speedup in my last message. However, if "we can simulate best parallel algorithm in a single processor without communication, not mention we usually have better sequential algorithms", assume compared to other speedup values, this is an unfair to use the method to get a speedup value. In my last message, the speedup that I said means the geneal speedup. The speedup that you used is called relative speedup in my monograph: Performance Evaluation, Prediction and Visualization of Parallel Systems, Kluwer Academic Publishers, March 1999. If possible, I would like to discuss it with you in detail. Xingfu Wu http://www.csc.lsu.edu/~wuxf Jun Zhang wrote: > By strict (original) definition, there should be no > superlinear convergence since we can simulate best > parallel algorithm in a single processor without > communication, not mention we usually have better > sequential algorithms. > > Some people go to conference to report results > of "superlinear convergence" to draw attentions. > In most cases, those results are obtained by > running the same algorithm on single and > multiple processors. (The algorithm is usually > the proposed algorithm.) So I think "superlinear > convergence" is actually misleading. > > Jun Zhang > ---------- >