Aucbvax.2440
fa.works
utzoo!decvax!ucbvax!works
Sat Jul 25 16:04:49 1981
Collected responses on global configurations
>From WorkS-REQUEST@MIT-AI Sat Jul 25 15:57:50 1981
------------------------------
Date: 24 Jul 1981 15:41 PDT
From: Deutsch at PARC-MAXC
Subject: Re: More on configuration
In-reply-to: BHYDE's message of 23 Jul 1981 1116-EDT
To: WorkS at MIT-AI
There is no "huge increase in cost of communications that
local networks imply versus front ends". The multiplexing
cost is the very large hardware (protection, etc.) and software
overhead required for time-sharing a mainframe. Ethernet-style
local networks have lots of bandwidth and are extremely cheap.
The bandwidth is needed to make reasonably fast access to shared
(either read-only or read-write) information feasible. If you
time-share a mainframe, the bottleneck becomes the processor
instead, and it's a lot more expensive to speed up the processor
than the communications network.
The cost of processor power is very misleading. If you are
talking about the power available to run a single thread (job),
then indeed there are great economies of scale. However, if you
want to run N jobs, the overheads of multiplexing processor and
memory, and the extra I/O required to get those jobs in and out
of the fast, expensive mainframe (which also goes up linearly
with N), may kill you. Proper planning for a computing facility
must make some judgments about the mix of computation-intensive
jobs, for which a fast, expensive, shared mainframe is necessary,
against the interaction-intensive jobs, for which workstations win
out. Similar comments apply to disks, printers, and long-distance
communications: whenever there is an economy of scale, you have to
decide how much of what capacity device you want. The one thing
where centralization produces DISeconomies of scale is interaction
bandwidth. That's what the increasingly cheap workstation tech-
nology is buying us.
------------------------------
Date: 24-Jul-81 11:38:50 PDT (Friday)
From: Hamilton.ES at PARC-MAXC
Subject: Re: More on configuration
Based on two years of living on an Ethernet, I maintain that any
network that does not give every user an effective data rate of
at least 56KBits/sec is of very limited usefulness. Even with 24
MBytes of local disk, we do a lot of file transfers -- swapping
files in and out if we're tight on space, retrieving new documents
and software, fetching obscure symbols files to track down bugs,
etc. You can get away with 9.6 KB phone lines for fetching/storing
on an overnight/ archival basis, but for real-time interaction, I
regard 56KB as the minimum acceptable.
The point about individual workstations being more cost-effective
has at least two major aspects: (1) Operating systems can be more
efficient when they don't have to worry about "fair" allocation
schemes and elaborate protection mechanisms. (2) System-wide
reliability is \far/ higher -- you never have hundreds of people
sitting around with the computer down, unless you have a massive
power outage. At worst, if the network goes down, people have to
wait a while to do printing or backup.
I've always thought Grosch's law, about computing power going
up as the square of the price, was a myth -- it's really more
like the square root of the price, when you consider the value
of redundant machines. Does anyone really believe that one
$6M 3033 system is worth more than 100 $60K PDP-11/24 systems?
There's also the phenomenon that software vendors tend to charge
in proportion to the cost of the hardware that their system is
to run on. The firstcopy of the same package might cost $500
for a CP/M version, $5K for an RSX-11M version, and $50K for
an MVS version.
--Bruce
------------------------------
Date: 25 July 1981 0946-EDT (Saturday)
From: Hank Walker at CMU-10A (C410DW60)
Subject: workstations vs big CPU
In general, the price/performance of a CPU is no longer a function
of scale (I am ignoring packaging). The rate of improvement of
fast expensive technology, such as ECL gate arrays, is no greater,
or even less than that of VLSI NMOS/CMOS technology. In other
words, the power per dollar is going up at the same rate in each
of these two price ranges. The point of communication costs can
be argued both ways. Ethernet costs something like $2/foot, plus
the transceiver cost of several hundred dollars today, soon to
come down. This seems more expensive than RS232 into DZ11's.
On the other hand, CMU is going away from its centralized front
end to terminal concentrators on the Ethernet. This is due to
the cost of stringing wires, and the difficulty in getting more
performance out of a centralized unit.
------------------------------
Date: 24 Jul 1981 1142-PDT
From: Michael Dolbec <CSD.DOLBEC AT SU-SCORE>
Subject: Re: More on configuration
To: BHYDE at BBNG
In-Reply-To: Your message of 23-Jul-81 0816-PDT
On the leverage of buying large systems:
I think you are essentially correct about leverage available
in buying a large system as opposed to many smaller personal
workstations but you miss another point. The problem is that
technology keeps blazing even after you buy your cost-effective
large system. You stay current for a small time and then you
start to fall behind current hardware and since the chunk of
money necessary to purchase your large system is substantial
you can't just go out and buy another right away. You must
wait a relatively long time before you can purchase your next
state-of-the-art mainframe [son of VAX or whatever].
Thus, your fast expensive technology rapidly gets outdated and
since you can't replace it immediately your fitting the concave
upward technology curve with a large step function (you buy a
large system and you touch the curve, as time progresses the
curve arcs up but you cannot afford to buy a new machine so
your curve goes horizontal, finally you buy a new machine and
touch the technolgy curve again).
If you choose the personal computer route you can fit the
technology curve much closer, with a much smaller step function.
The function represents the relatively small incremental cost
of buying new personal workstations every so often in order to
keep pace with the latest technology.
--Mike
------------------------------
End of collected responses on global configurations
***************************************************
-----------------------------------------------------------------
gopher://quux.org/ conversion by John Goerzen <
[email protected]>
of
http://communication.ucsd.edu/A-News/
This Usenet Oldnews Archive
article may be copied and distributed freely, provided:
1. There is no money collected for the text(s) of the articles.
2. The following notice remains appended to each copy:
The Usenet Oldnews Archive: Compilation Copyright (C) 1981, 1996
Bruce Jones, Henry Spencer, David Wiseman.
