Freenet Developers,
Hi. I'm reading this list avidly and may start contributing source patches
if I have the time. In the meanwhile, I thought that some of you might be
interested in an essay I wrote today that covers the history and possible
future of distributed services as part of a vision of where I see projects
like freenet going.
If this sort of post is wholly inappropriate to the list, I won't repeat
the error.
Yours,
david weekly
[ from http://david.weekly.org/writings/clientserver.php3 ]
Client As Server: The New Model
A new model is emerging from the Internet. It represents the culmination of
years of incremental evolution in the structure of the network and the
clients that feed upon it. It is based upon the same principles upon which
the Internet was founded. It is this: the client is the server.
--- darpanet
The Internet was created as a distributed network. Originally conceived as
the Defense Advanced Research Projects Agency's network (DARPANET), it was
to be able to withstand a nuclear attack from the Russians. There could be
no single point of failure in the system, and in this way it had to be
different from most other networks yet conceived.
People previously had grown to the notion that there must be one, central
arbiter that oversees all transactions on a network: a Mainframe. This
model has an obvious weakness: when the Mainframe goes down, the whole
system is unusable. Then again, if there is only a singular important point
of failure, you could pay some people a lot of money to sit there and fix
problems as soon as they happened (and hopefully to insure that the
problems never happen in the first place). Unfortunately, it's difficult to
do this with regards to a nuclear bomb. So a different model was needed.
DARPANET provided this model by removing the server. It's sort of like
taking a model where everyone hands their mail to the post office, having
people pass a letter to a friend, who passes it to their friend, who passes
it to the recipient. While at first this might seem a little odd or maybe
even a little inefficient, it means that it would be a lot harder for
someone to stop a flow of mail to you (or a flow of mail in general).
Instead of simply bombing the post office, now they've got to assassinate
each and every one of my friends to prevent me from getting mail. Going
back to the real world, there would be no single point of failure where the
Russians could bomb to take down our communications.
It was a revolutionary as strange way of thinking about things. To this
day, some people don't understand it and ask questions like "where is the
server that runs the Internet?" or even "where is the Internet?" It's hard
to understand that every server that is on the Internet is a part of the
Internet.
--- availability
These days, we are amidst an equally paradigmatic change which almost
perfectly mirrors the first. Corporate servers, which distribute
information and services to clients and participate in "e-business" need to
not crash. Companies like eBay whose computers crash often get a bad name
and lose billions of dollars on their company's valuations, almost a worse
fate than the actual millions of dollars in customer transactions that go
out the door when servers die.
A quick fix is to employ a large number of servers configured exactly the
same, such that if one goes down, traffic is quickly diverted to the
others. Work is equally distributed amongst these servers by use of a "load
balancer." This solves a few problems, but what if your server cluster is
in California and the network link from California to New Zealand is
getting bogged down? While the long-term answer is to invest in a faster
connection to New Zealand, the short-term way to solve this problem is to
put a server cluster in New Zealand. This sort of rapid expansion can
quickly get expensive to deploy and manage. Some bright kids from MIT
figured this out a few years ago and cobbled together what is now one of
the fastest growing companies out there: Akamai. (Hawaiian for "cool" if
you're wondering.)
Akamai has already gone through the trouble of buying several thousand
servers and putting them in network closets all around the world. The idea
being that you can spoon off delivery of the parts of your site that don't
change much (the pictures, the movies, etc.) to Akamai and they'll take
care of making sure that your readership can always quickly access your
content. Cute idea. "Cool," even.
Distributed services lead to higher data availability. The more machines
that are distributing your content in the more places, the more people will
be able to access your content quickly. It's a straightforward idea. This
notion of distributing work is also useful for distributing computation...
--- processing power
It's expensive to build really fast chips. It's really expensive. To make a
chip twice as fast as today's high-end consumer chips costs about ten times
as much money. That's largely because today's consumer chips are, pretty
much by definition, as fast as it is possible to make them and have them
still be reasonably cheap. If this wasn't the case, another company would
have come along and made reasonably cheap, screaming fast processors and
have swept the market away. But the tight competition in the chip
manufacturing business has kept the "bang for the buck" ratio screaming up
through the ceiling, much to the delight of consumers.
It's important to note that making fast chips is expensive, because if I
want ten times the processing power that comes in a top-of-the-line
consumer PC, the best way to do that and save your money is not to buy a
machine that's ten times faster, it's to buy ten top-of-the-line consumer
PCs. People have understood this general concept for a long, long time:
wire together a bunch of processors to get a very, very fast machine. It's
called "Massive MultiProcessing" (MMP) and is pretty much how all of the
supercomputers of yore (and of today!) work.
The recent concept is that it's possible to do this with off-the-shelf PCs.
Recently, software (such as Beowulf) has been developed to make it very
easy to make a cluster of PCs act like one very fast PC. Sites that
previously had deployed very expensive custom supercomputer systems are
actively investigating using massively distributed commodity hardware to
serve their computing needs. That would be remarkable as-is, but this
concept of distributing computing cycles has gone even further than clumps
of commodity hardware: it's gone into the home.
--- seti @ home
For roughly the last forty years, there has been a serious and
conscientious effort to search for intelligent life in the universe by
listening for patterns in radio transmissions. The process of analyzing
incoming radio transmissions for patterns such as those that might be
emitted by intelligent life forms is mind bogglingly difficult, requiring
vast numbers of computations. While privately funded, the Search for
ExtraTerrestrial Intelligence didn't have enough money to process all of
the data that they were getting. They did, however, have a sizeable fan
base. (A number of people on this planet think it would be pretty cool /
important to discover intelligent life out there.) So what did they do?
They distributed the work.
Some clever programmers put together the software used for analyzing the
data returned by the Arecibo antenna (the largest radio receiver on Earth),
put some pretty graphics on it, got it to act as a screensaver, and put it
on the web. Consequently, several hundred thousand people downloaded it and
ran it as their screensaver. While they're away from their computers, this
pretty screensaver crunches through vast quantities of data, searching for
patterns in the signals. The SETI project (as of this writing) in this way
has a "virtual computer" that is computing 13.5 trillion floating-point
operations per second, thanks to the people running the "screen saver."
Individual computers can be used to distribute mathematical work!
(I feel I should also mention distributed.net, which spends its time using
people's computing power to crack cryptography codes. Their "virtual
computer" is currently cracking a 64-bit cipher known as RC4 at the rate of
130 billion keys per second.)
--- data services
So it's now clear that it's advantageous to distribute computation and the
serving of data across as many computers as possible. We've seen how a few
projects have distributed computation across end users, but what projects
have distributed data services?
Napster is one of the first and best examples of end-users acting as
distributed servers. When you install Napster, it asks you where your MP3
files are. You tell it and it goes out and makes a list of what MP3 files
you have, how long each song is, and what quality the recording is. It then
uploads this list (but not the songs) to a central server. In this way, the
central server has a whole bunch of lists: it knows who has what music of
everybody who is running Napster. You can ask the server who has got songs
by Nirvana and then contact those other users (while your Beck tunes are
possibly getting served to some Scandinavian with a predilection for
American music). This model allows for information (in this case, MP3
files) to be rapidly and efficiently served to thousands of users.
The problem with it is both technical and legal: there is a single point of
failure: Napster's servers. While there is more than one server (the client
asks a "meta-server" what server it should connect to), they are all owned
by Napster. These servers unfortunately do not share their file lists
between each other and as a result, you can only share files (and see the
files of) others connected to the same server that you happen to have
connected to. Napster is currently being sued by the RIAA for acting as a
medium for distributing illegal MP3 files. While it is true that Napster
can be easily used for illegally distributing MP3 files, they themselves
don't actually copy the bits for users: a bit more like acting as a Kinko's
that happens to be used by subversives than actually making MP3 copies
themselves.
If you are a Napster user, you should be worried about this lawsuit,
because if the RIAA succeeds, they will probably want to come and shut down
Napster's servers, thus theoretically shutting down the whole Napster
network. In short order they could quickly close down any Napster clones
due to the legal precedent that the anti-Napster case would set. Boom. Game
over, no more illegal music.
Theoretically.
--- a virtual internet
The RIAA mentality is one and the same of the Russians of yesteryear: a
desire to stop the flow of information through the network. The answer to
the Russians is one and the same as the answer to the RIAA: a completely
distributed system. If every client on the network was connected to a
handful of other clients, each of which in turn connected to others like
some apocalyptically enormous online incarnation of Amway, then every
person could have some connection to other people through a chain of mutual
acquaintances. It's Six Degrees of Freedom. (There exists a theory that
says that on average you know someone who knows someone who knows someone
who knows someone who knows someone who knows anyone in the world. That is
to say, you are about six degrees from every human on the planet.)
This is a "virtual Internet" of sorts where links are not physical (a wire
from you to me) but logical (e.g., I know you). Data flows through this
"web of friendship" in such a way as it looks like you are only talking
with your friends, when really you are talking to your friends' friends,
and so forth.
--- gnutella
The same rebellious college hacker genius who created the fabulously
popular MP3 player, WinAMP (and was subsequently bought out by America
Online, now America
Online-Time/Warner-Netscape-EMI-And-Everything-Else-But-Microsoft) happily
hacked out a program that allows for the free exchange of just about any
kind of file over such a peered network. Unfortunately, his bosses
discovered it halfway through development and quietly tried to erase the
fact that the renegade project had ever existed in the first place. The
name of the program? Gnutella. (Named after the delicious chocolate spread,
Nutella.)
Since there's no central server around which Gnutella revolves, AOL's
shutdown of the project didn't actually stop Gnutella from working. A
growing user base of several thousand souls (myself included) uses the
product on a daily basis to share files of all types, from music to movies,
to programs. At last check, there were about 2200 people using it, sharing
1.5 terabytes of information. Wow.
There's no way to shut it down. There is no organization to sue to stop it.
There is no server to unplug that would bring the network tumbling down.
Because as long as at least two people are running the software, the
network is up and running.
--- freenet
There exist even more advanced projects in the works that will build upon
these notions to create an even more powerful incarnation of a peered
network that incorporates notions of perfect anonymity, trust, secrecy,
realtime communication, and even banking. Freenet is perhaps the furthest
along in this, although it has a very long way to go as of this writing. If
you're interested you can read about my own scheme for a Secure + Anonymous
File EXchange.
--- the future net
Akamai has shown that it is clearly advantageous to have content
distributed by as many nodes as possible: companies are willing to pay good
money to have their content on thousands of servers all over the world.
Gnutella is showing that it is possible to create distributed networks that
cannot be shutdown, even in the face of legal and technical opponents.
Napster shows that such networks can become popular and that people are
willing to act as servers. Seti at Home shows that people will even allow
others to use their computing power for a "greater good."
What is enabling this now? Well, computers are unsurprisingly getting
faster every year. The average desktop that's sold to Joe User for doing
word processing, email, and web browsing can, when properly configured,
deliver hundreds of thousands of email messages a day, serve millions of
web pages, route Internet traffic for tens of thousands of users, or serve
gigabytes of files a day. (Joe probably isn't aware of this and will still
kick it when Word takes five minutes to load.) His hard drive could store
100,000 websites each having ten or so pages, email for 1000 users and
still have room for a few thousand of his favorite songs. Furthermore, if
Joe has a DSL or a cable line to his house, he's got a static IP (an
address on the Internet that doesn't change often, if at all), is almost
always connected to the Internet, and is online at high speed.
In short, Average Joe's computer resembles one of the best Internet servers
of yesteryear.
If thousands of Joes end up running "community" applications like Gnutella,
they can take advantage of their connectivity, disk space, and computing
power. New "co-hosting" services will spring up like popcorn in the
microwave. Here are a few possibilities in that direction:
--- the distributed future
Visualize for a moment, sending out your website into a collective ether,
to be served by hosts around the world: if one computer goes down, others
will spring up to serve it. Your page never goes down. Your friends send
you email encrypted so only you can read it; it is stored on half a dozen
of your friends' computers, accessible to you from anywhere on the planet.
All of this in exchange for setting aside a small chunk of your hard drive
(100 megabytes or so) and a little bit of your bandwidth to serve web pages
and people's email. Any content that you consume (except for your personal
email!) instantly is rebroadcast over the network: your computer in this
way will help content flow to where it is popular.
--- requiem server
In the future, there will be no need for centralized services. All content
will be available on this peered network. Strategies for "partial
consumption" such as letting people read the first few paragraphs and
charging for the story or hearing the low-quality song and charging for the
audiophile version will be adopted along with anonymous payment schemes. It
will be possible to send out intelligent agents to this network to search
for books, music, or other merchandise: clients (such as Amazon.com, CDNow,
eBay, your neighbor, etc.) that have a match for the merchandise will
communicate with you through the peered network, preserving your anonymity.
You will be able to make an anonymous payment (or merely a secure payment
if you prefer) and your goods will be on their way to you. No more URLs. No
more servers that crash, email that is unavailable, websites that you can't
get to, or data that you can't find. It will truly be the end of the server
as the line between what it means to be a "client" and a "server" on the
network becoming increasingly blurred to the point of indistinguishability.
---------------------------------------------------------------------------
-----
a footnote on wireless
It is worth pointing out that wireless Internet access may well become
democratized as well. Since high-speed wireless Internet is taking so long
to reach America, citizens may well do it themselves. Apple recently
popularized the IEEE 802.11 standard for wireless Ethernet by including
AirPort in the iBook. Wireless Ethernet cards are now available for PC and
Mac desktops and portables. More exciting yet, people have been working on
extending the range of the AirPort from a couple hundred feet to tens of
miles. You can imagine now a future a few years away where one person every
city block has got a base station and everyone else jacks in. Some have
postulated this as being the ultimate incarnation of the communality and
free spirit of the Internet.
_______________________________________________
Freenet-dev mailing list
Freenet-dev at lists.sourceforge.net
http://lists.sourceforge.net/mailman/listinfo/freenet-dev
