Personally I'm in favour of stream based ones, but it depends how the traffic really looks as to what would work best. We have a gig link with 10 x 100mb tunnels across it. So *IF* there is traffic which needs to push above 100mbps then if it's stream based that's never going to happen, and as is often the way, a document with a bunch of signatures on it states per packet.
I can't see how per packet would be better, and i do have it going really well with standard observed lb, but ideally need at least a proof of concept if nothing else.
btw, is there any information on how / if the load balancing methods change when a pool is being used for routing as opposed to a vip? things like observed and predictive methods don't feel as applicable.