fyi

[Marshall Rose <mrose () dbc mtview ca us>]

Path: dbc!infoods.unu.edu
From: KLENSIN () infoods unu edu (John C Klensin)
Newsgroups: dbc.mrose.iesg
Subject: Re:  OSI v TCP
Message-ID: <754870793.898089.KLENSIN () INFOODS UNU EDU>
Date: Thu, 02 Dec 1993 17:19:53 -0500 (EST)
References: <9312021428.AA18307 () TIS COM>
Sender: daemon () dbc UUCP
Lines: 78
To: crocker () tis com
Cc: iesg () CNRI Reston VA US, iab () isi edu
X-Envelope-To: iesg () CNRI RESTON VA US
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; CHARSET=US-ASCII
Content-Transfer-Encoding: 7BIT
Mail-System-Version: <MultiNet-MM(330)+TOPSLIB(156)+PMDF(4.2)@INFOODS.UNU.EDU>


Having listened to today's comments on this, I decided to spend a few
minutes with my friendly stats package, just looking at the numbers.
The story they tell is pretty pathetic, almost independent of assumptions
about what the market really is and what is really being measured.  I
will speculate on the latter a bit, but it probably isn't worth much
trouble.


You can verify 90% of what follows in five minutes with a sheet of graph
paper.  I've fussed around a bit more than that; anal details on
request.


The projection method for the TCP "curve" is to assume a small amount of
growth between 93 and 94 (no clue where the number comes from), then
slightly less (but really about the same as 91-92 and 93-94) level 94-95
and 95-96.  This type of estimate shows less than linear growth over the
91-> 96  period.  I think it would be hard to observe any data that
would yield that conclusion, unless one assumes a rate of decline in the
_price_ of products far in excess of the rate of increase in product
use. 


In addition to the plot, you can get an interesting picture by taking
first differences--the annual growth amounts shown.  I.e., the TCP/IP
figures are:
  Year               91       92       93    94     95     96
  In the posting:   1.10     1.45    1.50   1.65   1.95   2.00
  Differences             .35    .05     .15    .30    .05


Note that this suggests an only 50 M $ growth in the TCP/IP market
during 93 after 350 M $ the previous year.  Those numbers ought to be
pretty easy to calibrate and are certainly not consistent with my
intuition.


Now look at the OSI figures.  91 to 93 is basically linear, within the
resolution of the data.  Then it shoots up.  Take logs and find the 
nicest little exponential growth curve anyone might make up.  The nature
of such a curve with this one's parameters is that a linear fit is about
good enough through 94.  And they are assuming that the OSI market is
larger than the TCP/IP one in 94.  That year starts in 28 days, folks--
I don't see the products out there to make that possible-- but let's
give them the benefit of the doubt for a second.  


In 95, it shoots up.  That is, of course, the same "OSI reaches critical
mass and really takes off two years from now" story that Dave commented
on and that I, too, have been hearing regularly for about a decade. 
Always two years from "now".  And the 94 figure reflects the other
oft-repeated fairy tale, that the OSI market will become larger that the
TCP/IP one "about a year from now".  Again, that story has been repeated
almost every year for a long time.


Normally, in order to project exponential growth, one must either have
exponential growth already (which we do with several measures of
Internet growth, but I've not heard argued for OSI) or some predictable 
event that would trigger such growth.  The only "event" that I can see
here is the "two year" story.


Conclusion:  These data are too good to be true, and probably don't
represent anything more than the same old mythology and wishful
thinking, adjusted to a 1993 starting point.  


Whether the values themselves, as distinct from their relationships,
mean anything, I hesitate to guess.  However, my own anecdotal
experience has been to see several large organizations, previously
firmly committed to OSI, retreat from it during the last 18 months. 
That can't be contributing to a positive trend.


It is also worth noting that "total sales of software in dollars"
figures are pretty easy to cook.   Most OSI products come as
separately-priced add-ons to other systems.  Much of the TCP/IP software
base, by contrast, comes bundled with the host systems.  One can make,
and justify, almost any assumption about, e.g., how much of the cost of
a UNIX system should be allocated to "TCP/IP software".  In the absence
of a good model, assumptions tend to be chosen that meet the needs and
agenda of the person/organization making the choice.


Fun stuff.   Anyone for a Seattle BOF on how to lie with network growth
statistics?   I don't think it quite fits in the ALE charter.


    --john