From: Wolf Dietrich Newsgroups: comp.sys.amiga.misc Subject: Re: ppc.library/WarpOS discussion Date: Thu, 16 Oct 1997 18:08:51 +0200 Organization: phase 5 digital products Lines: 51 Message-ID: <34463C13.578516F3@gf.phase5.de> References: <3442309A.74DF@mtek.chalmers.se> <32000306245443162499@BIRDLAND.es.bawue.de> NNTP-Posting-Host: ufoo.phase5.de Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: 7bit X-Mailer: Mozilla 3.0Gold (X11; I; Linux 2.0.0 i586) Path: 195.95.96.10!news.unisource.be!News.Amsterdam.UnisourceCS!cosy.sbg.ac.at!news.cs.utwente.nl!opus.ies-energy.com!news.webspan.net!newsfeed.internetmci.com!4.1.16.34!cpk-news-hub1.bbnplanet.com!news.bbnplanet.com!newsfeed.ecrc.net!newsfeed.nacamar.de!news.nacamar.de!not-for-mail Steffen Haeuser wrote: > > > But then again, coding like that does not work without losing lots of > performance. FORGET the independence tasks for each processor !!! Could you please explain on which qualification your statements are based? When you tell people to forget about the independent tasks on different processors, you tell them to forget about real parallel multiprocessing. You tell them to forget a solution that will feature accelerators with multiple PPC processors, which are already in the plannings and to be released next year. Why don't you talk about loosing performance in a kernel software concept that is completely focused on doing context-switches between CPUs? The asynchronous message system of the PowerUp System Software from phase 5 allows to send 1,000 (onethousand) messages (each with 30 byte of data enclosed) from a 68k task to a PPC task within less than 60 milliseconds (yes, for all 1,000 messages), and 1,000 replies from the PPC task to the 68k task in less than 8 milliseconds (these are actually measured values). Sending a message from one task to the other AND sending a reply needs less than 70 microseconds - THIS is a time that we consider to be fast. > CPU-inexpensive stuff you do on 68k side anyways... so you do not NEED > Multi-CPU for this. You really do not get any more speed out of it. Parallel > CPU would only bring you more speed for CPU-intensive (which also means > memory-intensive) stuff... but exactly THAT does not work without problems > that slow you down... well, i am telling this out of a programmer's point of > view... CPU-intensive does not always mean memory-intensive. When two (or more) CPUs share the same memory dynamically, as they do in the PowerUp design, they can consecutively read necessary data in burst accesses from the memory. Just a note from the hardware labs: When you watch the bus activity of the CPUs on a logic analyzer, you will recognize that even during application each of the CPUs is repeatedly off-bus, leaving access cycles for another bus master. When the second active CPU can use this time for memory access, the memory bandwidth is used much more efficiently, thus resulting in better overall performance. ------------------------------------------------------------ Wolf Dietrich, General Manager phase 5 digital products Return mail address: wd@gf.phase5.de http://www.phase5.de ------------------------------------------------------------