Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!umich!samsung!brutus.cs.uiuc.edu!psuvax1!psuvm!FERMAT!JEFF From: jeff@FERMAT.OTS.UTEXAS.EDU (Jeff Hayward) Newsgroups: bit.listserv.nodmgt-l Subject: BITNET Restructuring Proposal Message-ID: <9002100356.AA10746@fermat.ots.utexas.edu> Date: 10 Feb 90 03:56:18 GMT Sender: Node Management Discussion Reply-To: Node Management Discussion Lines: 42 Approved: NETNEWS@PSUVM Gateway Comments: To: NODMGT-L%BITNIC@pucc.PRINCETON.EDU In-Reply-To: "Michael R. Gettes"'s message of Fri, 9 Feb 90 11:09:22 EST <9002091706.AA09827@fermat.ots.utexas.edu> On Fri, 9 Feb 90 11:09:22 EST Michael R. Gettes said: > ... >Maybe I should explain some of the reasoning behind the selection of the >core sites. >First, considering the high volume of traffic going through >the existing core sites (those that participated in the BITNETII project >and to which most have been selected for the proposed core) at this time, >large scale machines were needed. We figured we had to find 3081 class >machines or higher. For example, PUNFSV2 (Princeton) is handling on >average 85,000 files per day at about 750 million bytes of data. Figures >from the NSFnet during the summer showed that BITNETII traffic accounted >for 25% of the NSFnet backbone traffic. We have 2 3081s here and we >use about 55% of one 3081 to support just the BITNET traffic that goes >through our site. So, horsepower is a significant factor. Also, as an Michael's fine elucidation of the logic behind the core structure struck a nerve, so I thought I'd generate some additional heat. It occurs to me, (and others, I think) that the wrong variables are being optimized here. Why should CREN build and perpetuate a structure that requires massive 3081 support, at large cost, to achieve an average 70k bits per second (750MB/day) throughput? I mean, really, a $4000 cisco router can do about three orders of magnitude better than that, 'cause that's what it's good at. It seems that if per-node connectivity, rather than per-link throughput, were optimized the whole network would be more efficient. Large scale NJE pipes wouldn't be necessary if the TCP-capable folks (that's most of us) could just talk to each other directly. We wouldn't need an NJE "core", or the techno-aristocracy that comes with it. The mainframes could go back to doing whatever it is they're good at, and the rest of us could just get on with business. Right? Don't get me wrong, VMNET is better than nothing, and much appreciated for its benefits. The core proposal should probably proceed (or has it already?), since there's nothing better at hand. I just think someone's made an error in their assessment of the fundamental problem. Well, that's my 2 cents. Just thought it ought to be said. Jeff