Hello Wang,
per packet load balancing increment jitter (delay variation) and inter-packet delay.
Some applications like VoIP can be negatively affected by this.
Another reason is that to implement per packet load-sharing a minimul state machine for each flow has to be held:
the device must remember when packet N+1 of a flow arrives out what interface has been sent out packet N or it needs to allocate a pointer to the list of links that is incremented at each packet.
CEF implements per-packet load-balancing efficiently but the side effects on jitter are still there
per destination load-balancing doesn't require a mobile pointer to the list of links: all packets of a flow are sent out the same interface.
This is more efficient
Hope to help
Giuseppe
