Path: utzoo!utgpu!watserv1!watmath!att!pacbell!pacbell.com!mips!zaphod.mps.ohio-state.edu!usc!cs.utexas.edu!uunet!crdgw1!crdos1!davidsen
From: davidsen@crdos1.crd.ge.COM (Wm E Davidsen Jr)
Newsgroups: comp.arch
Subject: Re: Re^2: Superlinear Speedup (was Re: Scalability?)
Message-ID: <2291@crdos1.crd.ge.COM>
Date: 4 Jun 90 14:53:12 GMT
References: <1990May3.203405.23456@ecn.purdue.edu> <2075@naucse.UUCP> <6897@odin.corp.sgi.com> <49622@lanl.gov> <1990May1.154558.24009@cs.rochester.edu> <1990Jun2.080658.12651@oracle.com> <1990Jun3.145408.2472@watmath.waterloo.edu>
Reply-To: davidsen@crdos1.crd.ge.com (bill davidsen)
Organization: GE Corp R&D Center, Schenectady NY
Lines: 21

In article <1990Jun3.145408.2472@watmath.waterloo.edu> sccowan@watmsg.uwaterloo.ca (S. Crispin Cowan) writes:

| It's a theorum that (theoretically, anyway) super-linear speedup cannot
| occur.  In practice, it may occur marginally, but this is due to the
| fact that P processors have:

  It fact and in theory there is a certain amount of overhead associated
with a multitasking system. This occurs even on an unloaded system. The
overhead can be expressed as the sum of the fixed cost (peripheral
interrupts, etc) and per-CPU costs (dispatching, some memory
management).

  In a well designed multiprocessor system the % of CPU in overhead goes
down as processors are added, since the fixed overhead is being split
between several CPUs. This continues as CPUs are added until some other
resource runs out, then the performace becomes sublinear again.
Typically memory bandwidth runs out, but other resources can become a
factor, too.
-- 
bill davidsen	(davidsen@crdos1.crd.GE.COM -or- uunet!crdgw1!crdos1!davidsen)
            "Stupidity, like virtue, is its own reward" -me