Do you note that the above problem identification specifies a *single**connection*? What part of "single connection" is unclear to you?
Do you understand that 'per connection' balancing does not work when: 1) the characteristics of the traffic are _unknown_ at the time of connection initiation, and 2) the characteristics of the traffic *change* _during_ that single connection.
The rest of the problem description that you chose not to quote gave a concrete example of the problem -- using a first user with a streaming audio stream, and a bunch of other people then doing something as simple as having multiple HTTP requests ("keepalive" protocol option) over a *single* connection. If you don't like that, consider 'passive' FTP -- where where a number of users first pull a small 'readme' file, and then (still within the _same_ connection), a multi-megabyte binary.
Without 'co-operation' from the other end of the links, the 'per-connection' approach _is_ the 'best' you can do. And in the average case, it does work FAIRLY well. In pathological cases, however, it can degrade to barely more than the capacity of -one- circuit. Those pathological cases are relatively rare, but they*do*exist*. And, when you get into one, performance gets =really= poor, for comparatively _long_ times.