From: Rick Summerhill, Associate Director of Backbone Network Infrastructure  <rrsum@internet2.edu>
Date: Mon, 17 Feb 2003 16:36:37 -0500 (EST)
To: abilene-ops-l@INDIANA.EDU
Subject: MTU sizes

At the recent Joint-Techs meeting in Miami, there was a lot of discussion about MTU sizes.  This is clearly an important issue in making sure large flows can occur across the network.  I wanted to make all of you aware of the recent discussions and of the recommendations that have come out.

For example, at the connectors BOF, we encouraged making 9000 bytes be the standard for all connectors to Abilene.  To increase the MTU size to Abilene, a connector should simply contact the Abilene NOC to  coordinate the appropriate changes.  In addition, it is equally important for connectors to raise the MTU size back through their own infrastructure, and not just through to their Abilene connection.  We encourage all connectors to increase their MTU sizes over the coming months.

Moreover, the JET has now recommended a peering standard that sets MTU sizes to 9000 bytes for all Federal networks.  Abilene will be in the process of coordinating this activity with all such peers.

Finally, at the MTU BoF in Miami, the following recommendations were made from an Internet2 perspective:

---

Internet2-wide Recommendation on IP MTU

Engineers thoughout all components of the extended Internet2 infrastructure, including its campus LANs, its gigaPoPs, its backbone(s), and exchange points, are encouraged to support, whereever practical, an IP MTU of 9000 bytes.

The rationale for this recommendation includes the following points:

• Applications, including but not limited to bulk TCP, benefit from being able to send 8K (i.e., 8 times 1024) bytes of payload plus various headers. An IP MTU of 9000 would satisfy this application need.
• A growing number of routers, switches, and host NICs support IP packets of at least 9000.
• Very few routers, switches, and host NICs support IP packets of more than 9500. Thus, there is comparitively little motivation for a value much more than 9000.
• There is anecdotal evidence that Path MTU discovery would be more reliable if a given agreed-on value were commonly used. This relates to weaknesses in current Path MTU discovery technology.
• 9000 is an easy number to remember.

It is stressed that this is an interim recommendation. Engineers are also encouraged to explore the benefits and practicalities of much larger MTUs, up to the full 64 KBytes permitted for an IPv4 datagram.