After 35 years of technology crusades, Bob Metcalfe rides off into the
sunset
I JOINED MY FIRST computer crusade as a systems programming student
at the Massachusetts Institute of Technology in 1965. There have been many
crusades since then.
In this final episode of From the Ether, I'll return to a crusade for which we
are now only stacking our lances: Anticiparallelism.
In the 1960s, punched-card batch-processing mainframes were defended by
IBM and the BUNCH (Burroughs, Univac, NCR, Control Data, and
Honeywell). I joined the encircling crusaders: Digital Equipment, Data
General, and Scientific Data Systems. Our siege machines were interactive
time-shared minicomputers.
In the 1970s, with teletypes on the rise, I defected to the personal computing
crusades at Xerox, later Apple and Sun, and eventually IBM, Intel, and
Microsoft.
In the 1980s, I led the Ethernet crusade. 3Com and Novell took up LANs
against stand-alone PCs and their defenders -- including Microsoft -- for
whom sneakernet and RS-232C were then quite enough.
And then in 1986 or so, Senator Al Gore "took the initiative in creating the
Internet," our current crusade -- led since by Vint Cerf, Sun, and Cisco.
Actually, the Internet has been its own series of crusades since 1969: remote
log-in, file transfer, e-mail, newsgroups, client/server, instant messaging,
Web publishing, push, Web commerce, streaming media, mobile,
peer-to-peer, and soon broadcasting Web entertainment.
Among my favorite current crusades is the all-optical Internet; with fibers
extending to the curb, to the home, and not soon enough, to the wall. Fiber
crusaders are thwarted for now by the powerful copper monopolies.
Another favorite is the Pay-As-We-Go Internet. Too many still think they'll
get for free, or for some low flat fee, all they desire in Internet access,
transport, services, and content.
And then there's the crusade for a secure Internet. Today's Internet is
anonymous to the bone. But anonymity must be the exception. And as soon
as we dethrone the encryption-controlling Clinton-Gores, communications
on the Internet should be strongly encrypted all the time.
Not to tilt at windmills, there's also Anticiparallelism, a crusade of mine you
probably don't recall from 1998. But I've just met someone actually working
on Anticiparallelism -- at Microsoft, no less.
Anticiparallelism is, in short, using idle resources to preprocess anticipated
computing tasks in parallel with current tasks. It's doing this automatically,
using pervasive, fine-grained, decision analysis in new applications, user
interfaces, programming languages, operating systems, and hardware.
Here are some nitty-gritty examples of problems that Anticiparallelism might
solve.
Start: What is your computer doing those first minutes after you start it up? It
has been idle for hours. It will be idle for much of the time thereafter. So why
doesn't it handle its duties before or after starting up?
Next: Why does linking to Web pages, especially the obvious next page,
take your computer so completely by surprise?
Back: When you hit your browser's back button, why doesn't your
computer have your previous page cached?
Reply: Why must your computer format and send your reply to the current
message before fetching and rendering your next in-box message? Why
aren't the sending of your last reply and the fetching of your next message
done in parallel with displaying your current message?
Disk: Why do we waste so much time waiting for disks to be backed up and
defragmented? With a little Anticiparallelism, both could be done in the
background as soon as a file is updated.
These are several of the most obvious problems that could be solved with
Anticiparallelism, but generally aren't. Today's computers squander
parallelism. Their software sequentializes computations, failing to anticipate
tasks that might be accelerated with precomputed partial results put in
inventory --cached -- using idle resources.
When I first wrote about Anticiparallelism, readers had issues. One was that
allowing a defragger to run in the background slows response. Another was
that background tasks too often crash PCs. True, Anticiparallelism requires
fine-grained pre-emption and five nines (not nine fives) of reliability.
Others complained that prefetching Web pages is immoral -- tragedy of the
Commons and all that. It's clever to invest our own idle resources on
potentially moot precomputations, but it's antisocial to do so with freely
shared community resources. Amen.
When I risked talking about Anticiparallelism this summer before the
Defense Advance Research Projects Agency, a funny thing happened. A
member of the prestigious Information Science and Technology (ISAT)
Study Group came forward to say that he is already working on it.
Eric Horvitz calls it Continual Computation. He leads a group at Microsoft
Research and was excited that someone shares his passion.
What do Horvitz and I have in common that would lead us to this crusade?
Decision theory. We both rejoice in techniques for making better decisions
under uncertainty.
Horvitz and I know, for example, that after you choose one of three doors
behind which are hidden two goats and a car, and after Monty Hall shows
you a goat behind one of the two unchosen doors, the strategy that doubles
your chances of winning the car is to switch from taking what's behind the
door you first chose to taking what's behind the unopened door.
Horvitz's work in Continual Computing ranges from the theoretical to the
practical. He proves theorems about precomputation policies that maximize
the "flux" of computation. He deals with uncertainties about which computing
"challenges" will arise. He considers when it's right to employ idle but also
shared resources to advance one's own Continual Computing.
We talked about control structures in programming languages and operating
systems that might enable Continual Anticiparallelism. The normal sequential
structures such as "Do this, then this, then this, then end" allow for very little
parallelism.
Programmers can pursue parallelism by hand. Horvitz points to Microsoft
Outlook, which does some precomputing already. Another example is SQL
Server 2000, which uses an analysis of database workload to precompute
partial answers to queries.
But Horvitz and I, as do most crusaders, have high ideals.
Case dispatches such as "Do this, or do this, or do this, depending on what
the user types" are seldom used to guide parallel precomputation among the
cases while users make up their minds. But they could be.
As a first step, programmer-assigned priorities could guide the contingent
precomputation. Variations on an "agenda" control structure could let
programmers say, "When you get a chance, do this and this and this, in
parallel with these priorities."
But Horvitz and I are crusading for even more continually anticiparallelistic
computing systems, which will pursue optimization opportunities
automatically. We are crusading for self-aware computing that monitors how
often this follows that, that zeroes in on precomputations with high values
and probabilities, that prioritizes them using optimizations from decision
theory, and that makes the most of Continual Anticiparallelism.
For more on Horvitz's crusade, see
www.research.microsoft.com/users/horvitz.
Until we meet again, stay enthusiastic.
Missed a column? Click here for more
Bob Metcalfe today ends his decade-long run at a Pulitzer Prize. After
writing here for eight years, Bob is often voted InfoWorld's favorite
technology pundit. So he's declaring victory and quitting while ahead. You
can still buy his book, Internet Collapses and Other InfoWorld Punditry,
now in its third paperback printing from IDG Books. Bob promises we have
not heard the last of him. Farewells to metcalfe@infoworld.com. |
