This article was previously published under Q233401
This article has been archived. It is offered "as is" and will no longer be updated.
When a UDP datagram is larger than the MTU size of physical mediaand there is no ARP entry for the host it is sent to, Microsoft Windows TCP/IP implementation keeps only the last fragment of the UDP datagram sent to a given destination while waiting for an ARP reply. The rest of fragments are silently discarded.
For example, when a WinSock application attempts to send a single UDP datagram with 12501 bytes of data,the IP layer performs fragmentation and generates nine IP fragments on an Ethernet. The first eight fragments are discarded and only the last fragment is kept while waiting for an ARP reply. When the first ARPreply is received, only the last fragment is sent.
This behavior is by design and complies to Host Requirement RFC stating that ARP should save at least one packet.
To avoid the UDP packet drop:
Send a smaller IP datagram than the MTU size before sending a large UDP datagram. This ensures the ARP cache is "primed" so that all IP fragments can be sent out without waiting for the ARP reply
Increase ARP cache life by adjusting the following registry: