Lessee if I can explain this.. There's a lot of ways to bond together a series of circuits to make a larger pipe. You mention a few, so let me hit them..
Inverse MUX (multiplex) is when you take a special CSU/DSU that can combine the data from multiple circuits (usually T1's) and deliver them to the router as a single, large pipe. This has many advantages - The router config is very simple, as it just thinks it's got a 3Mb/s device and doesn't know that the CSU is doing inverse muxing to combine two T1's. This is really OSI Layer 1 stuff. Because it's layer 1, it is transparent to the other layers. This means that you can use Inverse Muxes for connections to frame providers, point-to-point circuits, etc. If it'll run across a T1, you can run it across an IMUXED T1. The only catch is, of course, that the other end has to have the same (or compatible gear) to IMUX their end, too. The catch with this is, of course, that the router doesn't ever see anything beyond the CSU, so it doesn't have visibility into circuit status, etc. You could have 1 T1 down and it wouldn't know, other than reduced throughput.
IMA - Inverse Muxing for ATM is somewherere between layer 1 and layer 2. It's a standard way to split up a stream of ATM cells for distribution across several circuits, then recombining them at the other end. It's an ATM-only protocol. Cells are usually load balanced using a round-robin technique. This works OK for ATM, since ALL the cells ar e 53 bytes, but not well for other protocols, which have varying size frames. I wouldn't worry too much about IMA, as you need to have a pretty serious network to need it.
Multi-Link PPP (MLPPP) is another layer 2 protocol that takes two standard PPP circuits and "bonds" them together. It's like IMA for the masses, in a way, and is most typically used for ISDN, bonding the 2
"B"earer channels. Some use it for multi-modem use, too. PPP and MLPPP is used for point-to-point circuits only and won't support other WAN protocols like Frame Relay, ATM, etc. The important thing to understand is that MLPPP is a router-to-router protocol, combining traffic across two "normal" circuits, i.e., two T1's with separate normal CSU/DSU's.
OK, enough of the lesson.. As I mentioned previously, we use a lot of inverse muxes in my organization. They work pretty well, but are not perfect. You get a lot of redundancy, since about 50% of your outages are someone at the CO doing something stupid - Looping a circuit, unwrapping a wire, etc. and if you have multiple circuits muxed together, you can survive. Unfortunately, there is a "retraining" period, when a circuit goes down and comes back up. It generally doesn't involve the loss of connectivity, but if there's too much flapping of the circuit, it can take the whole thing down and require a reset.
There's always a price-point tradeoff for bundled T1's vs. a fractional DS3. Our main office happens to be 2 blocks away from the CO in downtown Seattle, so DS3's are pretty cheap for us, since they are charged by the FOOT for distance. The break even point is usually somewhere around 4-6T1's, depending on your LEC and local charges.
I'm a fan of bundling no more than 4 T1's. If you need more bandwidth than that, you probably need a fractional DS3, and should just bite the bullet. I've always seen a growth curve, where when you need 6Mb/s, you will usually push that to 8, then 10, the 12 over the next year or so. Better to be positioned for growth than be behind the 8 ball.
- G
