Path: news.mitre.org!blanket.mitre.org!philabs!newsjunkie.ans.net!newsfeeds.ans.net!news-was.dfn.de!news-spur1.maxwell.syr.edu!news.maxwell.syr.edu!rill.news.pipex.net!pipex!uunetukout!join.news.pipex.net!pipex!uunetukin!server1.netnews.ja.net!lyra.csx.cam.ac.uk!nmm1 From: nmm1@cus.cam.ac.uk (Nick Maclaren) Newsgroups: comp.arch Subject: Re: IA64 Self Virtualizable? Date: 21 Nov 1997 09:05:22 GMT Organization: University of Cambridge, England Lines: 47 Message-ID: <653isi$4od$1@lyra.csx.cam.ac.uk> References: <64q6l9$q0v@crl.crl.com> <64tj62$pm6$1@murrow.corp.sgi.com> <3474A736.36A6@boston.sgi.com> NNTP-Posting-Host: ursa.cus.cam.ac.uk In article , Anne & Lynn Wheeler wrote: > >the whole count-key-data out-board search was design point trade-off >regarding 64kbyte operating system memory and several hundred >kbyte/sec smart adapters. configurations were on the wrong side of the >trade-off by the mid-70s; system memory for caching disk locations was >by then cheaper than tieing the I/O subsystem with linear searches. Yes, quite. Cambridge went from fixed-block disk architectures to a 370/165 with MVT, and we were very rude about having to regress 20 years. >I could get 3* speed-up for most i/o instensive workloads with >fixed-block architecture and reasonably filesystem. I find that surprising, because I could get 95%+ of hardware transfer rate into an unmodified Fortran program under MVT with, admittedly, JCL that switched on chained scheduling (for BSAM, of course). AND I could do it for multiple transfers in and out of a single-threaded program! However, your later text (omitted) indicates that you are thinking more of searching and positioning, which were definitely grim. Simple sequential transfer went quite fast, until the MVS 'improvements' that broke chained scheduling and replaced by a much inferior (but much simpler) alternative. The interesting thing is that Unix is PRECISELY the converse. We are having severe trouble with I/O performance on many systems. Complex positioning, searching etc. are quite fast, but simple data transfers are GHASTLY. This is mainly because of Unix's model of synchronous, copying data transfer (cue the cache and context switch thread). Many years ago I claimed that Unix was the logical negation of MVS in the sense that, if either got something right, the other got it wrong! I have once found one minor aspect they both got right, but have now forgotten what it was. There are, of course, plenty of aspects that they both get wrong. Nick Maclaren, University of Cambridge Computer Laboratory, New Museums Site, Pembroke Street, Cambridge CB2 3QG, England. Email: nmm1@cam.ac.uk Tel.: +44 1223 334761 Fax: +44 1223 334679