Hello Alexandr,
these terms are general and not specific of EIGRP.
First routers were workstations with special software: each packet had to be "process switched" by the main cpu.
In developing router platforms Cisco and other vendors had introduced different generations of forwarding optimizasition tecniques: the first one to be developed was Fast Switching and the more modern and more used now is CEF (Cisco Express Forwarding).
Let's consider the communication between two hosts in two different subnets:
with process switching each single packet is processed in the same way using software and cpu resources.
However, all the packets after the first one could reuse the activity made for the first packet:
the router has found out what is the L3 next-hop for the packet (either another router or the destination end device) and it has found the way to build the frame (it has made ARP for mapping the next-hop ip address to the correct MAC address or the correct serial encapsulation)
This means that packets after the next one need only to be forwarded : the routing lookup has already be done.
This type of optimization traffic driven is called Fast switching.
CEF is topology driven: it builds all the information to forward packets before traffic is seen using routing table, ARP table and so on: with CEF also the first packet of a flow is just forwarded (switched) and doesn't require a routing lookup.
However, not all packets can be processed by CEF: some type of packets still need to be process switched: for examples routing protocol messages, CDP messages, STP BPDus and packets with some options in the IP header.
per-destination load-sharing means flow based load balancing: when multiple paths are available a flow with a specific source and a specific destination is fowarded in the same way: all packets use the same path.
per-packet load-sharing: the parallel links are used in round-robin a state information about what link has been used for packet N is used to decide out what interface to send packet N+1.
Fast switching uses normally destination based load balancing and it is not able to perform per packet sharing.
CEF uses both source and destination in flow based load balancing or it can perform per packet load balancing with an explicit configuration.
To be noted that usually flow based load balancing is to be preferred because it avoids out of order arrival of packets classified in the same flow
EIGRP specifics are that it can allow load-sharing over paths not of equal cost
Hope to help
Giuseppe
