Larry Roberts | A pioneer looks to balance Internet scale

Along with VintonCerf, Robert Kahnand Leonard Kleinrock,Larry Robertswas one of thefounders of ARPAnet,the network developedby the DefenseAdvanced ResearchProjects Agency thatlater became theInternet.Forty years afterRoberts began workingon ARPAnet, theInternet is beingoverloaded withvideo and peer-topeernetworking,which it wasn't designedfor, he said.Roberts also is thefounder and chairmanof Anagran, acompany that offersproducts to manageIP traffic. I had this perceptionthat we would makeknowledge available instantlyto everybody in the world. Itwas getting to be all on computers,and computers wereincompatible. And I thought,we need to move from languageto printing press tocomputer network. Now, theWeb is doing that, but we'redoing a lot more. The personalcommunications, thevideo, the other things, we'llget to, but basically, that wasn'tthe first thought. Before Idid the network, there wasthe knowledge network. All the computerswere incompatible.We had no way to move stuff.Even at [the MassachusettsInstitute of Technology] Ihad various computers atvarious places, and to movestuff between them, we hadto find a compatible media, atape that worked in all ofthem, which was not alwayseasy. So I started doingexperiments there, proved itworked, and we did it in anetwork after laying out thedesign. Ray Tomlinsonactually built the file transfermechanism called SendMessage and Read Message.He started sending me stuffbecause I was in charge ofthe program. And it was likea teletype that was pouringout. It wasn't very functional,because you couldn't reply,you couldn't look at the messagenames, you couldn'tstore, you couldn't forward.So I read the first e-mailenvelope in 1971. E-maillooks about the same today. I actually hada DEC minicomputer in myoffice that was the size of afiling cabinet. Most of thepeople on the network hadbigger computers. Everybodyhad different ones. When I leftthe Information ProcessingTechniques Office at DARPA,they finished TCP as a protocol.They got that on the networkin 1983. Then by 1986,we realized it needed to slowdown when things got lostbecause the network threwthings away if it was overloaded.From 1986 on, it wasstable, because TCP wouldbalance with the switchingequipment, which was justdumping random packets.Now dumping random packetshurts the voice, hurts thevideo and makes it almostunworkable. It doesn't workwith them, because they'refixed rate, so they won't slowdown. So it doesn't even helpto discard the packets. AndTCP will slow down, but italmost comes to a grindinghalt and stalls. It's a veryawkward system in responsetime, the way it's working.What I'm trying to do withFlow Manager is make achange in the way the networksundo the controls. Inother words, it's got to balancewith TCP.The latest thing we're seeingis that peer-to-peer isovercrowding the network.Basically, P2P is unfair, in thesense that if you have morethan one flow, you get N timesthe bandwidth, and everyflow gets equal bandwidth, soyou keep adding flows andyou do better than everybodyelse. That's unfair to everybodyelse who didn't use it.And so we've come to theconclusion that the averagerate ... should be the same for each home or for each user. Itdoesn't matter how many flowsyou use. So P2P can do itsthing, everybody can do theirthing, but they get fair use ofthe network. We don't look insidethe packets to do that, wedon't go searching and violatingprivacy, we just control therates of the flow. So that turnsout to be a radical improvement.It makes it fair, and it almoststops the net neutrality fight,because now you don't have toargue about it. We put fournodes in, starting in October1969. By the end of the year,we were on the West Coast:the UCLA, Stanford ResearchInstitute, the University ofCalifornia at Santa Barbaraand the University of Utah.UCLA was the first, SRI wasthe second, then we quicklyexpanded it nationwide in1970. By 1971, we did a bigdemo, which illustrated to allof the press and the publicthat it worked, and they wereshocked by it. The researchpeople had seen it, but theworld hadn't really. And thetelephone people who hadbeen saying it would neverwork finally realized that itwould. And they threw rottenapples at us. They couldn'tbelieve that this would work. It was text anddata. We transferred binarydata for photographs. We didalmost immediately start testingvoice. And voice had lowenough bandwidth that therewasn't a big problem, we knewthat would work. But we didn'tthink video would ever be feasible.We knew it would haveto be vastly faster, like 100times faster, like it is now. ...Nobody saw video coming. It's still notdesigned for video. It doesn'tscale. Changes will have tohappen to make video workover the network in the scalewe're using it. That's why wehave to potentially block otherusers, if there are 10,000 otherpeople using it. In five years, Icertainly hope we have a significantdeployment of systemsthat protect us and provideequality, as I said about peer topeer.Secondly, I would hope thatwe are managing video properly,so the kind of Flow Managerthings we're doing will help.They will get deployed over fiveyears, I think, by somebody.Thirdly, I don't know howwe're going to achieve it, butwe have to add network security.I want the network to tellme that I know that that's youI'm communicating with andyou've been verified. So youwould have to have then a certifiedID when you got on thenetwork. And then everythingyou did would be approved,and I'd know who you wereand I wouldn't get spammed.