The lost Van Jacobson paper that could save the Internet
rdist.root.org
rdist.root.org
A pivotal recognition is that the Internet nowadays is dominated by content, not by endpoints. The modern CDN is really a very complex kludge in order to deal with what we now believe the Internet's primary purpose to be: content delivery. The Internet was originally designed in a TCP/IP point-to-point structure, with the most important aspects being the communication of two parties. Now the most important part is that a given party (or really many parties) receive a specific piece of content, regardless of where that content comes from. In this light, it is wise to redesign the Internet's infrastructure against standard TCP/IP and instead build knowledge of content and caching layers directly into the transport protocols.
The preliminary paper which deals with many of these issues is Jacobson's Networking Named Content[1]. The general idea is that one requests content instead of communication with a specific party. ISPs and other intermediate layers cache content blocks (because its in their own financial best interests by saving money in peering costs and pure networking hardware costs) and your request for content is delivered via the nearest available hot cache. Many other papers have been written with their own tweaks, but this is the general idea for what the next generation of the Internet should look like (from the distributed systems researcher perspective). In my Distributed Systems in Challenging Environments class one of the things my partners and I were working on was building a prototype of CCN (the protocol described in Jacobson's paper) on top of the current TCP/IP layer to examine its real-world characteristics and swarm behavior. We never completely finished the prototype for Contemplating Content Centric Protocols, but it was an interesting project nonetheless.
[1] http://conferences.sigcomm.org/co-next/2009/papers/Jacobson....
I've been hoping to see something tangible from VJ on this for awhile now. He was doing content-based networking even back when he was at Cisco.
To an extent I think recognized by a lot of networking researchers, including IIRC YC's Robert Morris at MIT, the future of the Internet is probably not IPv6 and a network consisting of operating systems addressed by scalar integers.
Instead, IP-type protocols are going to assume the same role that Ethernet and 802.11 (and ATM) play today, as a substrate for connectivity on top of which real applications will be built. The wide area discovery and "session" (OSI style) management roles played by IP and TCP will be played by overlay networks. We have primitive overlays now (DHTs, BitTorrent, Skype) and proprietary ones (CDNs like Akamai, whatever the cable content providers are using to push VOD to the edge), but eventually we'll get a general-purpose open one and the game will be on.
I think this follows straightforwardly from Reed's end-to-end argument. Intelligence belongs to the edges, not the core of the network; intelligence at the core is necessarily a lowest-common-denominator affair, and is hard to scale. That's a key reason why we don't have multicast today. Routers are so bad at scale that tier 1 NSPs filtered BGP announcements by prefix length, because they couldn't handle /32-granular routing for even the tiny subset of machines that actually wanted it (begging the question of how we ever expect IPv6 to do the things the typical HN reader hopes they'll do, like provide them with their own portable /16-equivalents). Multicast effectively asked those same overtaxed routers to address web pages, individually.
That's obviously not going to work when the service model involves random people demanding that AT&T and Level3 add routing table entries for their podcasts. But it scales just fine when the thousands of machines interested in those podcasts share a protocol and an infrastructure for arranging a fan-out overlay.
So many unsolved problems here, from "what is the most reasonable kernel of overlay assembly, group management, and routing to deploy to end-systems" to "how do we avoid congestion collapse in arbitrary group messaging protocols" (a study of multicast reliability algorithms --- at least, from the late 90s when I last did this stuff --- will probably horrify you). But the great thing about it is, nobody has to ask permission to make this stuff work; we can evolve to the real next Internet without getting Comcast's permission, or for that matter (more perniciously) the IETF's.
The problem we were trying to address in the graduate class is that Jacobson's paper simply assumes that the Internet suddenly "switches," like a lightbulb, to this new method of routing. I think everyone agrees that something that looks like CCN will become the future, but we were interested in the road to get there. We were trying to examine the problems associated with porting to this kind of network without preexisting infrastructure in place.
One thing that you may find interesting is that the network tends to look more like a very layered Bittorrent network in current IP infrastructure, so it may end up that the most effective research ground for an intermediate protocol would be in adapting Bittorrent. The second area that I think is really interesting is secure and (perhaps more importantly) authenticated communication in this protocol. There's an entire other paper cited at the bottom of the CCN paper that I linked to that details the cryptographic authentication used in CCN but there are a number of things that I think either need more detail to be worked through or put me on the edge a little.
We had a centralized tree-structured directory for discovery, and then (at first) deployable nodes running group messaging and running a weighted link-state routing protocol (more or less cribbed from OSPF), then later a small kernel message forwarding scheme with a programmable control plane so we could build arbitrary routing protocols in Tcl.
Our initial application was chat (we overengineered, like, a little) and we pivoted to streaming video.
We died in part because we hadn't the slightest clue what we were doing, and in part because the VCs replaced our management with a team that decided our best bet was to take our superior technology and go head-to-head with Akamai with our own streaming video CDN.
We'd have been better off just open-sourcing.
Anyhow: I'm obviously biased, but the way I think this is going to happen is, some open source project to build arbitrary overlays is going to catch on (the overlay will be a means-to-an-end for some killer app people actually care about; my bet is, telepresence).
Instead, we were more interested with the robust swarm implementations that were already present in some Bittorrent clients, like fairly efficient DHTs. It is also an already existing protocol with millions of users so we could piggyback on Bittorrent clients with plugins to measure the real performance all over the world.
The key difference is the caching. Jacobson proposes that the caching would happen in the ISPs, but there's no reason peers couldn't take up the slack, only different monetary incentives. The protocol here becomes less like OSPF and more like distributed k-minimum spanning tree. If you can find the "supernodes" then you can implement caching on their front and prevent unnecessary communication.
Of course, if you think anything like me your first thought is, "do we actually know if data requests are reliably concentrated in a specific area such that local caching will actually provide much of a benefit?" My partners thought the same thing, which is why they ended up doing a Bittorrent traffic analysis to attempt to plot an AS distribution curve for data requests[1]. They found that there was huge redundancy in intra-AS requests, so there's at least potential here.
Of course there are still a number of problems to be worked out -- if the supernodes don't have an easy incentive to provide caching to other people in their AS, how do we develop such an incentive? But the overall protocol seems very interesting. Most notably it would not work like FEC and small files would probably benefit even more from the caching (which is why you'll also note that files are split into blocks or datagrams or whatever you wish to call them in the VJ protocol).
[1] http://acm.eecs.northwestern.edu/site_media/cccp.pdf
I really don't deserve to be on this paper -- Zach and John really ran this show.
On the other hand, congestion control is easier to do for FEC-based protocols than it is for generic group messaging.
One thing that makes FEC attractive in wide-scale group messaging is that you have knobs to turn, in terms of how many simultaneous blocks you attempt to download to reconstitute the message.
As you can imagine, this is an idea that works really well for ISOs and DVDRips, but is clunky for emails.
I've only just started working on understanding swift so I can't speak for its suitability for small content. It is much simpler than bittorrent at least.
https://gettys.wordpress.com/2010/12/17/red-in-a-different-l...
VJ is at PARC now.
The old VJ LBL papers are still collected at ICSI/ICIR (Vern Paxson is still there!):
http://www.icsi.berkeley.edu/cgi-bin/pubs/browse.pl?groupid=...
There used to be some attention paid to deliberate evasion of congestion avoidance (Stefan Savage did a paper on it), the idea being that a malicious user could monopolize a link by ignoring congestion signals (interview question: when you upload a file via Wi-Fi over your DSL connection, how does your computer know how fast to send packets?). Maybe some of the reason that turned out not to matter was, we were in a temporary period of not needing congestion avoidance.
Finally, as I understand it, at least some of the fundamental ideas behind VJ's "hydrodynamic" congestion avoidance scheme are derived from Raj Jain's CUTE, which was done for DEC in '85.
The Stefan Savage paper you refer to wasn't so much about ignoring congestion signals as deliberately faking reception of data that hadn't actually been received. With things like duplicate acks or optimistic acking, you could cause the sender to open their window size to whatever you want.
http://cseweb.ucsd.edu/~savage/papers/CCR99.pdf
Back in 1999, I was working on integrating these ideas into a router product. It would be software in your home gateway that would keep a cache of the last observed window size for each pair of hosts for a few seconds.
If you started another TCP connection to the same host within that short interval, the router would spoof both sides to increase the window size to the previous value / 2 and then let additive increase take over from there. This bypasses slow start but the current available bandwidth is likely to be very close to the estimate within that short range and hosts didn't cache window size, even when making multiple connections in a short period (e.g., HTTP 1.0)
I archived this paper and a bunch more here. Prepare to lose an afternoon if the history of TCP/IP interests you.
Also there's a huge problem on routing algorithms due to hundreds of patents covering very trivial uses. For example, there is an awful patent covering the very important case of not dropping tiny ACK packets. For congestion control, it's more useful to drop big TCP data packets to make the endpoints adjust. And the key networking gear players are full on supporting those crazy software patents.
So my prediction is things will stay the same and nothing meaningful will happen for 10 years until a few key patents expire. Nobody will risk exposure to a multimillion IP lawsuit. And anyway, we the users will just pay the bills either way so telcos don't care.
You can thank Big Telcos, Cisco and many others.
Of course big players with a huge patent war-chest (i.e. Google/Amazon/Microsoft) will advance their WANs.
I built and sold a product that cheated on TCP in order to minimize negative customer experiences during periods of congestion.
It was a one line change in the linux networking code, and suddenly their crappy networks were no longer a problem!
(Just so you don't think me completely evil…
All traffic was across their own networks.
Their networking folks considered 5% packet loss to be acceptable.
The product's purpose was to aggregate and tunnel connections. Multiple TCP inside TCP gets unfairly punished by the congestion algorithms, which is why most VPNs don't use TCP, but sometimes you find yourself under the thumb of administrative rules and you have to do it.
I simply capped our back off time to something like 10 seconds. Short enough that no one thinks the device broke because "the network was stuck, but now I can ping but your device is hung so I'll power cycle it". If your network can't take 1 packet per 10 seconds on behalf of a few dozen aggregated connections, it isn't me that broke your network.)
http://cseweb.ucsd.edu/~savage/papers/CCR99.pdf
(And Nate's right, I drastically oversimplified my summary of it). This is "cheating TCP" by Stefan Savage from '99. A great, great paper.
RED was designed at a time when implementing packet schedulers was expensive (theory was not completely there yet, firewalls were not widespread, etc.). Now there are fast packet scheduler implementations available. Our own QFQ [1] is distributed with the official FreeBSD and Linux sources, and runs reasonably fast even on an OpenWRT device (remember that AQM needs to act at the bottleneck [2], which is mostly your DSL router for upstream traffic).
You could really try it out easily, and if you are willing to set up a few dummynet pipes [3] (FreeBSD/Linux/Windows) you don't even need to act on the DSL router.
References:
[1] http://info.iet.unipi.it/~luigi/qfq/ but the code is already part of recent FreeBSD and Linux distributions (edited: and even on OSX you'll find an older but similar scheduler as part of ipfw+dummynet), you just need to load the relevant modules.
[2] there were, long ago, some proposals/products which did ack pacing on the reverse path to do cong.control on the forward path. I think it is tricky to make them work.
[3] http://info.iet.unipi.it/~luigi/dummynet/ . You can set two pipes with a bandwdith slightly below that of the DSL link, so you can control the bottlenecks. How can you know what is the bottleneck bandwidth is a different story. You could look at the negotiated speed on the modem, often available through the web pages, but probably you'll have other bandwidth caps elsewhere. You could try to run some of the BW probing apps to some nearby host, hoping that the bottleneck is your local DSL, and dynamically reconfigure the pipes accordingly. I have not tried this yet.
The Click codebase is one of my favorite pieces of C++ code ever.
SOPA if it passes may create such pressure. It's main target is DNS.
CDN's like Akamai rely on universal use of one DNS, the one that SOPA aims to regulate, to accomplish their kludge.
Food for thought.
I prefer "Chicago" to any of the other alternatives.
End-to-end can be realised. Overlays work for small groups. Small groups can share with each other. Networks connecting to other networks. It's been done. ;)
Evolutionary pressures may help.
Generally one of the reasons I don't freak out about laws regulating the Internet is that the Internet as we know it today is rickety anyways.
In the last 10 years we've seen the monumental shift (predicted in the late '90s but widely scoffed at) from ad hoc network protocols and native client/server implementations to a common web platform. Nerds still recoil a little at this, thinking about the native/ad-hoc stuff they still use (like Skype), but if you look at the industry as a whole, particularly in the line-of-business software development that actually dominates the field, it's all web now. TCP ports are mostly irrelevant. If you were going to start a new app today, would it be a native implementation of a custom protocol? Probably not!
One of the things that got us to this state was in fact regulatory: default firewall configurations that don't allow anything but web out, and disallow arbitrary inbound connections.
Over the next 10-15 years, I'm hoping we'll get similar nudges away from first-class addressing only for machines (which are less and less useful as an abstraction for completing tasks using technology) and towards first-class addressing for subjects, interests, content, &c. This is an insight lots of people have had, from David Cheriton & TIBCO in the '90s through the RON work at MIT through VJ's work at PARC & so on.
I wrote off Lessig for a bunch of years after reading _Code_, but I think he fundamentally has it right in the sense that implementors have as powerful a say in how things are going to be regulated as legislators do. America had the Constitutional Convention after the Articles stopped scaling; the Internet will have little blips of architectural reconsideration when the impedance between the technology and what people want to legitimately do with the technology gets too high.
(I'm trying to make a technical point here, not a political one; I support copyright, and am ambivalent about SOPA.)
In situations where anycast is used, how do you even know what machine a given address to connecting to?
RON was a step in the right direction, imo. With small overlays, MAC addressing comes into play and it becomes a little easier to know what machines (not simply what "addresses") you are really connected to.
As long as it's an answering machine and the message is the same at every telephone reached by this number, it does not matter.
But as soon as you want to reach a live person, and not simply "content", then what?
Is end-to-end about accessing "content" or is it about communicating with someone operating another computer?
One opinion is that a very important Internet application will inevitably be 1-1.
Who did the FCC just hire as their new CTO? What is happening to POTS?
1-to-many systems, hacked to give an illusion of 1-1 conversations, e.g. smtp middlemen, social networking's http servers or twitter-like broadcast sms, are what we settle for today, but, imo, this is a limitation not a desired goal.
The last statement is true only for TCP/Reno. Reno here refers to a congestion detection algorithm, and it is that of looking for packet loss to recognize the congestion. This is the simplest and oldest approach, and it has been suceeded by TCP/Vegas, which looks at delivery delays instead, or how it is called a congestion avoidance algorithm. Basically the idea is that if the route gets congested, there will be buffeting happening at some hop and that would cause the RTT along the route to climb up.
This being said, I don't know the split between Reno and Vegas adoption in the real world. However I think newer Windows versions are on Vegas.
ftp://ftp.ee.lbl.gov/email/vanj.94mar14.txt
http://root.org/ip-development/e2e/end19940314-diagram.ps
In summary, RTT variance as a congestion signal is actually quite bad. It's normal to have variation even without congestion, so utilization would be worse if you backed off every time RTT increased.
> It's normal to have variation even without congestion
Certainly, but the question is if congestion-based variation is substantially different from an ambient variation. I am guessing that it is. Reacting to just a RTT increase is pretty dumb. The congestion marker clearly needs to be a bit more advanced than that.
I suspect what is happening is that the uptake of video services like YouTube and Netflix has resulted in link capacity being exceeded. The result is not pretty. The link acts like it is dead for about a minute at a time. It periodically starts working for a few seconds, maybe up to a minute, but then dies again. I suspect this is the oscillation effect at work.
If it were not for a draft that was discovered, this paper would have been lost forever.
I used RED on an OpenBSD firewall years ago at a small ISP that had too small of a pipe for the number of sites they were hosting out of their closet. It didn't seem to make a remarkable difference, but I wasn't responsible enough at the time to do careful metrics before-and-after.
So to update our detection scheme, we need routers to get feedback about the >queueing<.
Google+ existed as early as 1988?