What you are seeing is normal. It's very hard to get a 'true' load-balancing amongst all links. It's not a per-packet load-balancing.

A src-dst flow may hash via one link for 'conversation A' and hash via another link for 'conversation B'. 'Conversation A' may require more bandwidth and will show on the % utilization on the interface.

Please check out this article:

http://www.cisco.com/warp/public/473/4.html

Good points here. It's a tradeoff for quick hardware hashing, you will never get perfect load balancing.

You can, however, change the etherchannel load balance criteria. Depending on what type of traffic this is, you will see better results with different inputs.

For instance, load-sharing on MAC addresses when you have an L2 etherchannel between 2 routers, you will get poor balancing since there are only 2 mac addresses in that network. If your traffic has multiple flows per IP address, you will likely get better results by using src-dst IP and L4 ports:

"port-channel load-balance src-dst-port"

12.2(33)SXH also has some enhancements in this area.

If you are curious to calculate which link a given flow should be taking, you can use:

"rem login sw"

then

"test etherchannel load-balance interface

address pair>"