[lir-wg] Discussion about RIPE-261
Michel Py michel at arneill-py.sacramento.ca.us
Wed Jun 25 05:51:07 CEST 2003
Aleksi, Agree with your posting below, which is one of the reasons vendors of geeky hardware have stated that they can't guarantee they'll continue to ride Moore's law (one of the other reasons being that the problem of routing table growth tends to switch from memory/cpu usage to protocol stability issues). > These technologies are also way more complex than DRAM > to begin with. No argument here. Put a DRAM wafer and a CAM wafer side-by-side under a microscope and you'll immediately realize this. Michel. > Aleksi Suhonen wrote: Routers -- and especially L3-switch-routers -- tend to use content addressable memory (CAM) or other such specialized architectures instead of dynamic random access memory (DRAM) to store the forwarding table on NICs. The production numbers of DRAM chips are orders of magnitude larger than those of CAM chips, and the gap is getting bigger. These technologies are also way more complex than DRAM to begin with. This means that the same amount of memory takes up a lot more silicon real estate which is exponentially relative to chip cost. Using just DRAM is not an option either because you cannot build the whole Internet on small-to-medium size routers with slow line cards. Some operators will always outgrow them and those operators will be the ones who in the end dictate how things are done. This doesn't actually have anything to do with my personal view on routing table growth. I'm just saying that there is a real technical reason to be worried about it. CAMs (et al.) are getting larger every year of course but, to the best of my knowledge, they are not doubling in size every year as per Moore's law. To quote Frederick Brooks, I don't think there will be a silver bullet to this problem.
[ lir-wg Archives ]