On Mon, Apr 19, 2010 at 7:43 AM, Edward Ned Harvey <lop...@nedharvey.com> wrote: >> From: Tom Limoncelli [mailto:t...@whatexit.org] >> Sent: Monday, April 19, 2010 7:29 AM >> To: Edward Ned Harvey >> Cc: da...@lang.hm; discuss@lopsa.org >> Subject: Re: [lopsa-discuss] IPv6 and NAT >> >> On Mon, Apr 19, 2010 at 7:03 AM, Edward Ned Harvey >> <lop...@nedharvey.com> wrote: >> >> From: discuss-boun...@lopsa.org [mailto:discuss-boun...@lopsa.org] >> On >> >> Behalf Of da...@lang.hm >> >> >> >> I know that the people pushing IPv6 consider NAT evil and want to >> >> make it impossible. >> > >> > Reference please? I never heard anything like that. I would agree >> that >> > it's generally not expected to be encouraged. But made impossible? >> I call >> > BS. >> >> There are certainly people in the IETF that are anti-NAT. They've >> been making their case for quite some time, including >> http://www.faqs.org/rfcs/rfc1627.html (which I agree with). >> Are they making it difficult for people to use NAT with IPv6? Yes. >> When people come up with a reason to do NAT+IPv6, they run to make >> sure the need is fixed some other way. For example, >> http://www.faqs.org/rfcs/rfc3041.html >> Are they making it impossible? Well, it depends on where "difficult" >> stops and "impossible" begins. > > This is not an attempt to make NAT impossible, or even difficult in IPv6. > There's no such thing as even making it difficult. If you want to build a > router, or a proxy server, and take internal connections and proxy them with > another IP address, go right ahead. Nothing's preventing you. Except for > maybe lack of customer demand.
That strategy does reduce customer demand. I think we're in agreement as to the limits of this technique. >> common multiplier. On the other hand... Subnets come in powers of two >> and racks hold 40 or 80 machines. The nearest power of two is 64 or >> 128. Both give you a 65% efficiency. Subtract out other > > So ... if I've got 80 machines in a rack and I've got 128 Ip addresses, I > can't use the remaining ~40 IP addresses in the next rack? That's a good suggestion on paper, but it becomes too difficult to manage in reality. At that scale, every rack has to look the same or things get unmanageable. The current architecture you'll see at most massive clusters is "1 rack = 1 subnet" or "2 racks = 1 subnet" (40 or 80 machines). Each rack (or pair) has a ethernet switch. All ethernet switches have uplinks to large routers. The economics work out best that way. Check papers published by Google, IBM and others for examples. Repairs, management, and power becomes easier because you can think in terms of racks as a unit. (at the risk of overstating my point: among the many benefits, you now don't have to buy patch panels and 2x the patch cables; at warehouse scale computing saving $40 per rack is real money; Saying "let's spend an extra $500,000 on patch cables so I can save some IP addresses" wouldn't get you much love.) And now the plug... I'll be giving a talk about some Google technologies at the LOPSA PICC conference. It's only a 45 minute talk, but I will mention some of these economics. Since MD is a short drive, I hope you can make it, Ed. PICC conference http://picconf.org May 7-8, 2010 Tom -- http://EverythingSysadmin.com -- http://www.TomOnTime.com Computer and network administrators... Spread the word! LOPSA New Jersey Professional IT Community Conference New Brunswick, NJ, May 7-8, 2010 -- http://picconf.org _______________________________________________ Discuss mailing list Discuss@lopsa.org http://lopsa.org/cgi-bin/mailman/listinfo/discuss This list provided by the League of Professional System Administrators http://lopsa.org/
