MTU negotiation

[rebelobrian]

My question is how hosts agree on a common MTU and if they do? Let’s say that one computer MTU is 1500 bytes and the other is 9000 bytes on an Ethernet network. Do the hosts negotiate a common MTU or a layer 3 device is needed to fragment the packets? I have heard from some people that a layer 3 device only fragments packets from different network architecture, ex: Ethernet and Token, not Ethernet to Ethernet. Other say that a layer 3 is necessary to fragment packets and other say that hosts will negotiate a common MTU so no fragmentation is needed. Which one is it? If someone feels confident that they know the exact answer, please enlighten me on this matter. BTW this is IPv4.

Thanks

Brian

[Richard Burts]

The question you are asking is complex and it is difficult to find any single answer that is right for each part of it. The biggest part of your question deals with negotiating MTU. The primary vehicle for negotiating MTU is Path MTU Discovery. In PMTUD an end station sends a frame at its maximum MTU and sets the bit that indicates do not fragment. As the frame is forwarded to the destination is it gets to a segment where it needs to be fragmented the device that needs to fragment (but can not fragment because the Do Not Fragment bit is on) should send an ICMP error message to the source indicating that Fragmentation was Required but DF set. The source station would recognize that a smaller frame was required and send smaller frames till it found a size that did not generate errors.

There is a major challenge in using PMTUD in many of todays networks because many networks have put policies in place that deny ICMP messages (especially ICMP generated from some other network). If the ICMP message can not get back to the source then the source will never know that its frames are too large and will not try smaller frames.

HTH

Rick

[rebelobrian]

Thanks for the input. I'm part of a team that maintains some networks that have workstations using Ethernet jumbo frames. As of right now the workstations have two IPs, one for 1500 MTU and the other for 9026 MTU. This is on the same NIC. We are trying to figure out if this is the most efficient way to deal with different MTU sizes. Another possibility was to place these workstations on a different VLAN and only use one IP, but we weren’t sure how MTU path discovery worked. If you have any more inputs, you are welcome.

Brian