Path: utzoo!mnetor!uunet!lll-winken!lll-tis!ames!mailrus!tut.cis.ohio-state.edu!bloom-beacon!oberon!cse.usc.edu!yueh
From: yueh@cse.usc.edu (Alan Yueh)
Newsgroups: comp.arch
Subject: Re: RAM Question:
Message-ID: <8737@oberon.USC.EDU>
Date: 26 Apr 88 17:01:35 GMT
References: <1005@iitmax.UUCP> <76700023@uiucdcsp>
Sender: news@oberon.USC.EDU
Reply-To: yueh@cse.usc.edu (Alan Yueh)
Organization: University of Southern California, Los Angeles, CA
Lines: 27
Keywords: Processors, wait states
Summary: Fractional wait states are being used presently

In article <76700023@uiucdcsp> gillies@uiucdcsp.cs.uiuc.edu writes:
>
>There should be a more precise metric for machines with caches that
>achieve zero wait-states.  For instance, if your machine is
>
>Wait State		When			How often	
>0			in the cache		97% of the time
>2			not in the cache	3% of the time
>
>Then I submit you are using a .06 wait-state machine.  The expected
>value of the number of wait-states per memory reference is .06. 

This metric is currently being used.  Both the Compaq 386/20 and the IBM
PS2 Model 80-111 (also 20Mhz) have a rated wait state of 0.8 (or is it
0.6?  I forgot.)  I'm quite sure that IBM releases the fractional figure,
I don't know if Compaq does as well.  However, lest I get flamed for
mistaking the PS2 structure, let me say the Model 80 does not get the 
fractional wait state figure from cache miss/hit ratios: rather it is a
function of accessing the memory within the same 2K page.  (OK, so it
looks like a cache without the separation of hardware.)

However, the publishing of the fractional wait states begs the question:
what's the benchmark that would reference a common 2K page?  I suppose that's
like the benchmark that determines a cache hit ratio.  Anyway, I don't know
the benchmarks that IBM and Compaq used to ascertain the wait states.

Regards, A.W.Yueh (alan)
	 (soon to be "formerly yueh@cse.usc.edu")