Anytime you can get divergent routing, go for it; however, if you're using the same ISP for both then you're going to want to know what you're using the T1s for: video/voice (real-time) or large packet transfers.
There are a couple load-balancing methods available to you, and these have changed in recent code versions. The old way was per-destination load balancing only. The problem here was cacheing could allow un-even load distributions if one conversation was typically more load-intensive than the next. Then came per-packet load balancing. This also had cacheing issues, which is why they backed off to process switching for this method. Your problem here was out of sequence packet issues which cause problems with voice and video. Then came Multilnk PPP (I don't even tackle Multilink, multichassis PPP).
MLPPP created a logical virtual-access interface that would make your two T1s look like a single 3 MB Full-duplex pipe. I've encountered two problems here. First, despite the fact that Cisco is employing PPP, a multiprotocol encapsultation, their support of IPX is clugey. You end up having to use Loopack interfaces (yes, a logical ontop of a logical interface) with commands such as ipx-ppp client. Very wierd. You also get to enjoy such caveats as Virtual-intefaces that never go away and are sometimes never used again, for seemingly no reason -- thus consuming memory and indice space.
Then Cisco came up with Cisco Express Forwarding. CEF deploys a router table-like and router cache-like implementation utilizing FIBs (forwarding information b...something (blanking on the moniker)) and another table. This is supposed to address some of the packet and destination load-balancing issues previously experienced. Thus, by using CEF load balancing, you should be able to load balance both T1s without having to worry about logical interfaces used by MLPPP implementations. My understanding, is that you'll still be subject to uneven balancing if you go destination load balancing since it's still a conversation to one T1 balancing.
Lately (as of 12.05T, though if youcan help it, don't use T codes (free warning)), Cisco has made improvements to MLPPP such that the virtual-access interfaces (whichyou really didn't have control over if you tried more than one on a box) are now replaced with multilink interfaces (similar to their multichassis MLPPP). This seems to be a better implementation, however, if you don't go the 12.05T, you'll have to upgrade to 12.1 LD codes, which require more memory.
As far as processor loads, I haven't seen huge changes in the loading. This, of course, dependent on the amount of process switched traffic.
