Interface Fragment table overflow. Input or output?

[netone.cisco]

Hi all. I've been getting the following message

%IP_VFR-4-FRAG_TABLE_OVERFLOW: GigabitEthernet0/2: the fragment table has reached its maximum threshold 64

Its now at 64 but has been at 16 and 32 and it grows as our clients grow. I do know why this happens but my question is:

Packets stay at an interface fragment table on input before being processed? My logic says yes but I would like to hear some confirmation and what you have to say.

Attached is a quick drawing of our edge network. Gi0/3 is a subnet of http caches where packets are forwarded by WCCP on a GRE tunnel. packets FROM the caches are not GRE encapsulated.

I hope my question is clear

Thanks to all

[Rodney Dunn]

They are held waiting for the subsequent fragments in an effort to reassemble all of them to get a complete view of the packet.

See:

http://www.cisco.com/en/US/docs/ios/12_3t/12_3t8/feature/guide/gt_vfrag.html

It's best to try and prevent fragmentation if at all possible. For example, if you know it's TCP and there will be some overhead due to a GRE header you could have the device adjust the TCP MSS size to account for it:

http://www.cisco.com/en/US/tech/tk827/tk369/technologies_white_paper09186a00800d6979.shtml