From: Peter Rander Newsgroups: comp.lang.fortran,comp.parallel.pvm Subject: Re: superlinear speedup, (was: fortran77 programming) Date: Mon, 14 Dec 1998 10:18:30 -0500 Organization: School of Computer Science, Carnegie Mellon University Message-Id: <36752C46.41C6@cs.cmu.edu> References: <3660D892.E1EDBF12@est.it> <3665044D.A0F57E59@kings.uq.edu.au> <3665E491.7993355@flash.net> <366BF6F3.1E8DC934@cic-mail.lanl.gov> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit Xref: ukc comp.lang.fortran:62158 comp.parallel.pvm:7869 Like many readers of these two bboards, I suspect, I have been sporadically following this discussion, waiting for the active players involved to discover that they are talking at two *EXTREMELY* different levels. On the side of the one-fast-processor-beats-N-slower(by_1/N)- processors, we have a *PURELY THEORETICAL* argument based on the assumption that NOT ONLY does the processor run N times faster, but that there are N times the number of registers, N times as much L1/L2 cache, N times as much memory bandwidth, N times as much... I think you get the idea: N times as much of *everything*. In addition, since this is theory, we can *STATE* that the processor does not need to conform to modern processor design. With N program counters, we don't need to push/pop state, etc, since we can just keep all those registers around. In fact, if this were actually done, then the argument goes that things might actually go *faster*, since process A might have an L2 cache miss, forcing a memory read, while process B then gets to process, which could be a computation -- thus reducing the loss of a main memory access. On the other side, we have numerous this-is-the-real-world- so-what's-your-point authors stating all kinds of things like "that would hurt modern processors", "you can't build such a system", etc. The arguments here are pretty clear, mostly valid, and almost obvious in many cases, even to a guy like me who is not a computer architect. So perhaps this thread can be split into "theoretical computer science issues in PVM and FORTRAN" and "real-world issues in PVM and FORTRAN" (PVM and FORTRAN being the two bboard homes of this incredibly bizarre miscommunication). I suspect that the discussion would have ended millenia ago with such a simple segmentation.... -Pete ----- Harvey J. Stein wrote: > > "Tony T. Warnock" writes: > > > A simple argument (due to Vance Faber of LANL) shows that > > superlinear speedup is not possible. Given an N times faster > > processor, one can have the processor execute the first instruction > > of each of the N slower processors, then the second instruction, > > etc. The faster processor will finish at least as fast as the > > slower set of processors. > > One can't do this and it'd make a modern processor slower. > > One can't do it because the the registers will be in the wrong state > > Even if you could do it, it'd make a modern processor slower because > you'd break pipeline optimizations that had been done to the > individual programs. > > -- > Harvey J. Stein > BFM Financial Research > hjstein@bfr.co.il