welty@steinmetz.ge.com (richard welty) (07/03/88)
In article <1988Jul1.043049.2418@ziebmef.uucp> becker@ziebmef.UUCP (Bruce Becker) writes: * it seems like the time to look at the practice of other systems - * in specific, I am familiar with BitNet, which send out descriptions * of available binaries (source, documents, etc), and issues a pointer * to a thing called a "listserver"... * It seems to me that this uses far less net bandwidth than the * broadcasting method, and serves the community equally well... In article <3335@s.cc.purdue.edu> rsk@s.cc.purdue.edu.UUCP (Rich Kulawiec) writes: >An extreme example, but more complicated topologies and cost distributions >eventually lead one to the same conclusion: if enough people at enough >sites want the source, it's cheaper to post it. Some years ago, I think >Chuq did some analysis of this problem, and concluded that the tradeoff >was somewhere around 100 people, in terms of overall network bandwidth. Chuq has stated that his analysis is now probably somewhat out of date, and recently guessed that the number was likely to be more like 200 (as I recall -- I did not save his article, unfortunately. >Clearly, however, this is a huge lose for the originating site, which >must send 100 copies of something rather than 1. It can be a huge lossage for links adjacent to the originating site as well. When I ran the auto-sports mailing list off of steinmetz, more than 40 of the addresses were being directed by pathalias through the phone link between rochester and steinmetz, which was pretty expensive. We attempted to configure sendmail here on steinmetz to fold addresses, but the copy that we have here (on an ultrix system) does not support the appropriate options. Of course, with a listserver, folding the addresses (if your sendmail can do it) would require batching requests, because otherwise the folding would not be possible. -- Richard Welty 518-387-6346 GE R&D, K1-5C39, Niskayuna, New York welty@ge-crd.ARPA {uunet,philabs,rochester}!steinmetz!welty ``I always wondered what that switch did'' -- Aaron Heller
chuq@plaid.Sun.COM (Chuq Von Rospach) (07/07/88)
>* It seems to me that this uses far less net bandwidth than the >* broadcasting method, and serves the community equally well... >>An extreme example, but more complicated topologies and cost distributions >>eventually lead one to the same conclusion: if enough people at enough >>sites want the source, it's cheaper to post it. Some years ago, I think >>Chuq did some analysis of this problem, and concluded that the tradeoff >>was somewhere around 100 people, in terms of overall network bandwidth. >Chuq has stated that his analysis is now probably somewhat out of date, >and recently guessed that the number was likely to be more like 200 >(as I recall -- I did not save his article, unfortunately. Whatever numbers I did are hopelessly out of date these days. With the proliferation of Telebit modems and PC Pursuit and the essentially free NNTP links slogging data around the country, the only way to really judge "cost" is to split the links into "free", "low cost" [PC Pursuit, local calls] and "high cost" connections. I haven't really looked at it, but it'd be extremely hard to even get a rough estimate of what percentage of the links fall into what category. Without that, trying to estimate costs and the E-mail/net crossover is essentially impossible. The limiting factors become, in many cases, disk space, CPU cycles, or modem access and bandwidth. When you're using NNTP or PC-Pursuit, the cost is fixed regardless of amount of data, your cost/byte goes to essentially zero. (I'm over-simplifying to some degree, here, but I think it's close enough for right now). You then get into a scenario where it's actually cheaper to send via USENET, even if only a (relatively) few people read it, and even if it goes on the exact same route -- because it's possible that the news is being held for PC-Pursuit during off hours and mail, on the exact same link, goes out immediately. So my numbers are useless in measuring reality today. If you want to try to build a "cost" of a given message via USENET versus E-mail and mailing lists, here's what I think you'd need to figure out the cost of a USENET message. If you take the number of sites on the net (random number: 7,000), and figure out the average number of hops needed to service each site (completely random number: 10), then take the percentage of "high cost" hops in an average path (random number: 40%), you can then cost of a message in money-hops (based on above random numbers: 7,000*10*0.40=28,000 money-hops). Factor in some overhead for the following: CPU Cycles, Disk space, modem utilization, fixed overhead costs for PC-Pursuit and/or Internet charges and administrative overhead. Turn that number into money-hops, just to keep thing consistent. As a completely random number, figure your fixed costs to be about 10% of your money-hops, so the total money-hop figure would be about 31,000). To turn this into a dollar figure, you'd have to come up with a cost per money-hop. Figure out the average cost of transmitting an fixed sized 1,000 byte message across a long-distance link @1200 baud (random number: $.02). So the (random) cost per K of a message on USENET would be: 31,000*$0.02, or about $620.00. (hmm. That message is compressed, not ascii. You can figure in a compression quotient). A 50K Macintosh binary would therefore cost the net about $31,000 in transmission charges. Expensive toys. These numbers are bogus. If someone's interested, they can try to figure out real versions of them. For instance, I'm willing to bet that the "high cost" number is exceptionally low -- I'd doubt that, on a net-wide basis, the "high cost" links account for less than 60% of the net, maybe as high as 75%. So these numbers may be as much as 50% low. Also, you have to remember to factor areas without full feeds (like Europe) in or out based on whether they get the newsgroups or not, so you can't just take a number like 7,000 for all cases -- cut out Europe and Australia and a given group might lose 10% of the potential sites, making the numbers worthless. To find the break-even point for a mailing list, you'd have to do the same kind of cost analysis as above. That's much easier (for the most part) because you know how many people are reading it and how many hops it takes to deliver the mailing list -- smart mailers in the middle rerouting on you being one major exception. The number of "high cost" links will be somewhat higher, because many sites don't hold up mail for PC-Pursuit distribution. As long as the money-hops are lower, it's cheaper to go by mailing list, except that you're also concentrating the strain on local machines, so there might be a regional bandwidth problem even though there's a net gain for the entire net. That can be a significant problem for some folks which might warrant dealing with the situation even if it's not otherwise justifiable. >>Clearly, however, this is a huge lose for the originating site, which >>must send 100 copies of something rather than 1. Yeah, but it really depends. 100 messages that go to the Internet is all CPU cycle cost. 100 messages that go to Australia via a modem is expensive. So you can't generalize just on number of messages -- you need to look at where they're going and how. Anyone want to comment on this analysis? It looks reasonable to me, understanding that the numbers I chose are bogus (kids, don't try this at home...), but these things always work better when multiple people smooth out the rough spots.... Chuq Von Rospach chuq@sun.COM Delphi: CHUQ Robert A. Heinlein: 1907-1988. He will never truly die as long as we read his words and speak his name. Rest in Peace.