In a press I'm often meeting the reports about the new emerging switching technologies, may be not quite new, as I understand those have been around for 2-3 years so far. As I could grasp the idea, it's Ethernet-based mesh network, utilizing Fibre Channel/InfiniBand features. It is my understanding that each switch fabric is capable to be joined in a meshed network, which allows to distribute data more efficiently, sort of computer clustering mixed with stackable switches.
I've heard of at least 3 companies doing this: Raptor Networks, Woven Systems, Arastra.
My understanding may be wrong of course. Can anybody tell more about this or point at good informative links?
I have no direct klnowledge of this. However, it seems like an attempt to replace the spanning tree used in Ethernet mesh networks with a mesh that looks more like what FibreChannel, Infiniband, or for that matter, IP routers, create among themselves.
Each node becomes the root of its own spanning tree, in essence.
i havent used these kit versions, but this is maybe the 5th or 6th iteration of a fix looking for a problem for L2 network resilience / load balancing.
So there are some lessons you can take from what has happened with earlier iterations of these types of "fix" in networking.
this is about fixing a problem that doesnt exist (or did exist but has been resolved already), and allowing big sprawling L2 networks without using spanning tree, and allowing traffic to cross multiple paths.
i dont have a problem with getting rid of spanning tree (or at least not depending on it to keep a network stable) - it is one of those protocols that works well most of the time, but when it goes wrong it tends to fail catastrophically.
ironically building a partial mesh of paths and pushing traffic down the optimal one is what routing protocols were designed for and what they are reasonably good at.
the standards based fix for the stated problem is to use some L3 switches in your campus and to dice it up so the L2 sections are relatively small, then let the routing handle the multipl paths and load sharing.
Once you do that the L2 "lumps" of network can be reduced in size - the logical minimum scope for a L2 domain is for a L2 domain is 1 switch.
for example a press story describes the Raptor switches as putting L2 routing info in "unused header space in Ethernet frames".
Last time i looked there wasnt any unused space in an Ethernet frame, which may mean they have tinkered with the frame format or "borrowed" some particular bytes - and that implies interworking with other kit may introduce all sorts of unpleasant surprises when you try it.......
So - you buy kit from this wonderful new 1 manufacturer with their special protocol, and you use it.
things that can go wrong:
you cannot get their compatible kit anymore (they go broke, get bought, or just decide they wont make the kit anymore).
you cannot get your maintainer to carry spares / find / train engineers on it.
you have to join their network to another one.
you have a bug and you need to capture some traffic to show the problem - Sniffer etc will not decode it....
1 golden rule of networking is - do not use a wierd, proprietary and / or poorly documented protocol or proprietary fix unless you have no choice.
if you cannot find exact technical descriptions of how it works on their web site on a public access page - i would run, not walk to another solution.
FWIW the only "enhancement" like this i have seen that i am comfortable with was the Nortel