IP MTU Discovery Options

STATUS OF THIS MEMO

A pair of IP options that can be used to learn the minimum MTU of a path through an internet is described, along with its possible uses. This is a proposal for an Experimental protocol. Distribution of this memo is unlimited.

INTRODUCTION

Although the Internet Protocol allows gateways to fragment packets that are too large to forward, fragmentation is not always desirable. It can lead to poor performance or even total communication failure in circumstances that are surprisingly common. (For a thorough discussion of this issue, see [1]).

A datagram will be fragmented if it is larger than the Maximum Transmission Unit (MTU) of some network along the path it follows. In order to avoid fragmentation, a host sending an IP datagram must ensure that the datagram is no larger than the Minimum MTU (MINMTU) over the entire path.

It has long been recognized that the methods for discovering the MINMTU of an IP internetwork path are inadequate. The methods currently available fall into two categories: (1) choosing small MTUs to avoid fragmentation or (2) using additional probe packets to discover when fragmentation will occur. Both methods have problems.

Choosing MTUs requires a balance between network utilization (which requires the use of the largest possible datagram) and fragmentation avoidance (which in the absence of knowledge about the network path encourages the use of small, and thus too many, datagrams). Any choice for the MTU size, without information from the network, is likely to either fail to properly utilize the network or fail to avoid fragmentation.

Probe packets have the problem of burdening the network with unnecessary packets. And because network paths often change during the lifetime of a TCP connection, probe packets will have to be sent on a regular basis to detect any changes in the effective MINMTU.

Implementors sometimes mistake the TCP MSS option as a mechanism for learning the network MINMTU. In fact, the MSS option is only a mechanism for learning about buffering capabilities at the two TCP peers. Separate provisions must be made to learn the IP MINMTU.

In this memo, we propose two new IP options that, when used in conjunction will permit two peers to determine the MINMTU of the paths between them. In this scheme, one option is used to determine the lowest MTU in a path; the second option is used to convey this MTU back to the sender (possibly in the IP datagram containing the transport acknowledgement to the datagram which contained the MTU discovery option).

OPTION FORMATS

Probe MTU Option (Number 11)

      Format
      
              +--------+--------+--------+--------+
              |00001011|00000100|   2 octet value |
              +--------+--------+--------+--------+
      
      Definition

This option always contains the lowest MTU of all the networks that have been traversed so far by the datagram.

A host that sends this option must initialize the value field to be the MTU of the directly-connected network. If the host is multi-homed, this should be for the first-hop network.

Each gateway that receives a datagram containing this option must compare the MTU field with the MTUs of the inbound and outbound links for the datagram. If either MTU is lower than the value in the MTU field of the option, the option value should be set to the lower MTU. (Note that gateways conforming to RFC-1009 may not know either the inbound interface or the outbound interface at the time that IP options are processed. Accordingly, support for this option may require major gateway software changes).

Any host receiving a datagram containing this option should confirm that value of the MTU field of the option is less than or equal to that of the inbound link, and if necessary, reduce the

MTU field value, before processing the option.

If the receiving host is not able to accept datagrams as large as specified by the value of the MTU field of the option, then it should reduce the MTU field to the size of the largest datagram it can accept.

Reply MTU Option (Number 12)

      Format
      
              +--------+--------+--------+--------+
              |00001100|00000100|   2 octet value |
              +--------+--------+--------+--------+
      
      Definition

This option is used to return the value learned from a Probe MTU option to the sender of the Probe MTU option.

RELATION TO TCP MSS

Note that there are two superficially similar problems in choosing the size of a datagram. First, there is the restriction [2] that a host not send a datagram larger than 576 octets unless it has assurance that the destination is prepared to accept a larger datagram. Second, the sending host should not send a datagram larger than MINMTU, in order to avoid fragmentation. The datagram size should normally be the minimum of these two lower bounds.

In the past, the TCP MSS option [3] has been used to avoid sending packets larger than the destination can accept. Unfortunately, this is not the most general mechanism; it is not available to other transport layers, and it cannot determine the MINMTU (because gateways do not parse TCP options).

Because the MINMTU returned by a probe cannot be larger than the maximum datagram size that the destination can accept, this IP option could, in theory, supplant the use of the TCP MSS option, providing an economy of mechanism. (Note however, that some researchers believe that the value of the TCP MSS is distinct from the path's MINMTU. The MSS is the upper limit of the data size that the peer will accept, while the MINMTU represents a statement about the data size supported by the path).

Note that a failure to observe the MINMTU restriction is not normally fatal; fragmentation will occur, but this is supposed to work. A failure to observe the TCP MSS option, however, could be fatal because it might lead to datagrams that can never be accepted by the destination. Therefore, unless and until the Probe MTU option is universally implemented, at least by hosts, the TCP MSS option must be used as well.

IMPLEMENTATION APPROACHES

Who Sends the Option

There are at least two ways to implement the MTU discovery scheme. One method makes the transport layer responsible for MTU discovery; the other method makes the IP layer responsible for MTU discovery. A host system should support one of the two schemes.

   Transport Discovery

In the transport case, the transport layer can include the Probe MTU option in an outbound datagram. When a datagram containing the Probe MTU option is received, the option must be passed up to the receiving transport layer, which should then acknowledge the Probe with a Reply MTU option in the next return datagram. Note that because the options are placed on unreliable datagrams, the original sender will have to resend Probes (possibly once per window of data) until it receives a Reply option. Also note that the Reply MTU option may be returned on an IP datagram for a different transport protocol from which it was sent (e.g., TCP generated the probe but the Reply was received on a UDP datagram).

   IP Discovery

A better scheme is to put MTU discovery into the IP layer, using control mechanisms in the routing cache. Whenever an IP datagram is sent, the IP layer checks in the routing cache to see if a Probe or Reply MTU option needs to be inserted in the datagram. Whenever a datagram containing either option is received, the information in those options is placed in the routing cache.

The basic working of the protocol is somewhat complex. We trace it here through one round-trip. Implementors should realize that there may be cases where both options are contained in one datagram. For the purposes of this exposition, the sender of the probe is called the Probe-Sender and the receiver, Probe-Receiver.

When the IP layer is asked to send a Probe MTU option (see the section below on when to probe), it makes some record in the routing cache that indicates the next IP datagram to Probe- Receiver should contain the Probe MTU option.

When the next IP datagram to Probe-Receiver is sent, the Probe MTU option is inserted. The IP layer in Probe-Sender should continue to send an occasional Probe MTU in subsequent datagrams until a Reply MTU option is received. It is strongly recommended that the Probe MTU not be sent in all datagrams but only at such a rate that, on average, one Probe MTU will be sent per round-trip interval. (Another way of saying this is that we would hope that only one datagram in a transport protocol window worth of data has the Probe MTU option set). This mechanism might be implemented by sending every Nth packet, or, in those implementations where the round-trip time estimate to the destination is cached with the route, once every estimated RTT.

When a Probe MTU option is received by Probe-Receiver, the receiving IP should place the value of this option in the next datagram it sends back to Probe-Sender. The value is then discarded. In other words, each Probe MTU option causes the Reply MTU option to be placed in one return datagram.

When Probe-Sender receives the Reply MTU option, it should check the value of the option against the current MINMTU estimate in the routing cache. If the option value is lower, it becomes the new MINMTU estimate. If the option value is higher, Probe-Sender should be more conservative about changing the MINMTU estimate. If a route is flapping, the MINMTU may change frequently. In such situations, keeping the smallest MINMTU of various routes in use is preferred. As a result, a higher MINMTU estimate should only be accepted after a lower estimate has been permitted to "age" a bit. In other words, if the probe value is higher than the estimated MINMTU, only update the estimate if the estimate is several seconds old or more. Finally, whenever the Probe-Sender receives a Reply MTU option, it should stop retransmitting probes to Probe-Receiver.

A few additional issues complicate this discussion.

One problem is setting the default MINMTU when no Reply MTU options have been received. We recommend the use of the minimum of the supported IP datagram size (576 octets) and the connected network MTU for destinations not on the local connected network, and the connected network MTU for hosts on the connected network.

The MINMTU information, while kept by the Internet layer, is in fact, only of interest to the transport and higher layers. Accordingly, the Internet layer must keep the transport layer informed of the current value of the estimated MINMTU. Furthermore, minimal transport protocols, such as UDP, must be prepared to pass this information up to the transport protocol user.

It is expected that there will be a transition period during which some hosts support this option and some do not. As a result, hosts should stop sending Probe MTU options and refuse to send any further options if it does not receive either a Probe MTU option or Reply MTU option from the remote system after a certain number of Probe MTU options have been sent. In short, if Probe-Sender has sent several probes but has gotten no indication that Probe- Receiver supports MTU probing, then Probe-Sender should assume that Probe-Receiver does not support probes. (Obviously, if Probe-Sender later receives a probe option from Probe-Receiver, it should revise its opinion.)

Implementations should not assume that routes to the same destination that have a different TOS have the same estimated MINMTU. We recommend that the MTU be probed separately for each TOS.

Respecting the TCP MSS

One issue concerning TCP MSS is that it is usually negotiated assuming an IP header that contains no options. If the transport layer is sending maximum size segments, it may not leave space for IP to fit the options into the datagram. Thus, insertion of the Probe MTU or Reply MTU option may violate the MSS restriction. Because, unlike other IP options, the MTU options can be inserted without the knowledge of the transport layer, the implementor must carefully consider the implications of adding options to an IP datagram.

One approach is to reserve 4 bytes from the MINMTU reported to the transport layer; this will allow the IP layer to insert at least one MTU option in every datagram (it can compare the size of the outgoing datagram with the MINMTU stored in the route cache to see how much room there actually is). This is simple to implement, but does waste a little bandwidth in the normal case.

Another approach is to provide a means for the IP layer to notify the transport layer that space must be reserved for sending an option; the transport layer would then make a forthcoming segment somewhat smaller than usual.

When a Probe Can Be Sent

A system that receives a Probe MTU option should always respond with a Reply MTU option, unless the probe was sent to an IP or LAN broadcast address.

A Probe MTU option should be sent in any of the following situations:

(1) The MINMTU for the path is not yet known;

(2) A received datagram suffers a fragmentation re-assembly

timeout. (This is a strong hint the path has changed; send a probe to the datagram's source);

(3) An ICMP Time Exceeded/Fragmentation Reassembly Timeout is

received (this is the only message we will get that indicates fragmentation occurred along the network path);

(4) The transport layer requests it.

Implementations may also wish to periodically probe a path, even if there is no indication that fragmentation is occurring. This practice is perfectly reasonable; if fragmentation and reassembly is working perfectly, the sender may never get any indication that the path MINMTU has changed unless a probe is sent. We recommend, however, that implementations send such periodic probes sparingly. Once every few minutes, or once every few hundred datagrams is probably sufficient.

There are also some scenarios in which the Probe MTU should not be sent, even though there may be some indication of an MINMTU change:

(1) Probes should not be sent in response to the receipt of

a probe option. Although the fact that the remote peer is probing indicates that the MINMTU may have changed, sending a probe in response to a probe causes a continuous exchange of probe options.

(2) Probes must not be sent in response to fragmented

datagrams except when the fragmentation reassembly of the datagram fails. The problem in this case is that the receiver has no mechanism for informing the remote peer that fragmentation has occurred, unless fragmentation reassembly fails (in which case an ICMP message is sent). Thus, a peer may use the wrong MTU for some time before discovering a problem. If we probe on fragmented datagrams, we may probe, unnecessarily, for some time until the remote peer corrects its MTU.

(3) For compatibility with hosts that do not implement the

option, no Probe MTU Option should be sent more than ten times without receiving a Reply MTU Option or a

Probe MTU Option from the remote peer. Peers which ignore probes and do not send probes must be treated as not supporting probes.

(4) Probes should not be sent to an IP or LAN broadcast

address.

(5) We recommend that Probe MTUs not be sent to other hosts

on the directly-connected network, but that this feature be configurable. There are situations (for example, when Proxy ARP is in use) where it may be difficult to determine which systems are on the directly-connected network. In this case, probing may make sense.

SAMPLE IMPLEMENTATION SKETCH

We present here a somewhat more concrete description of how an IP- layer implementation of MTU probing might be designed.

First, the routing cache entries are enhanced to store seven additional values:

MINMTU:

The current MINMTU of the path.

ProbeRetry:

A timestamp indicating when the next probe should be sent.

LastDecreased:

A timestamp showing when the MTU was last decreased.

ProbeReply:

A bit indicating a Reply MTU option should be sent.

ReplyMTU:

The value to go in the Reply MTU option.

SupportsProbes:

A bit indicating that the remote peer can deal with probes (always defaults to 1=true).

ConsecutiveProbes:

The number of probes sent without the receipt of a Probe MTU or Reply MTU option.

There are also several configuration parameters; these should be configurable by appropriate network management software; the values we suggest are "reasonable":

Default_MINMTU:

The default value for the MINMTU field of the routing cache entry, to be used when the real MINMTU is unknown. Recommended value: 576.

Max_ConsecutiveProbs:

The maximum number of probes to send before assuming that the destination does not support the probe option. Recommended value: 10.

ProbeRetryTime:

The time (in seconds) to wait before retrying an unanswered probe. Recommended value: 60 seconds, or 2*RTT if the the RTT is available to the IP layer.

ReprobeInterval:

The time to wait before sending a probe after receiving a successful Reply MTU, in order to detect increases in the route's MINMTU. Recommended value: 5 times the ProbeRetryTime.

IncreaseInterval:

The time to wait before increasing the MINMTU after the value has been decreased, to prevent flapping. Recommended value: same as ProbeRetryTime.

When a new route is entered into the routing cache, the initial values should be set as follows:

      MINMTU = Default_MINMTU

ProbeRetry = Current Time

LastDecreased = Current Time - IncreaseInterval

      ProbeReply = false
      
      SupportsProbes = true
      
      ConsecutiveProbes = 0

This initialization is done before attempting to send the first packet along this route, so that the first packet will contain a Probe MTU option.

Whenever the IP layer sends a datagram on this route it checks the SupportsProbes bit to see if the remote system supports probing. If the SupportsProbes bit is set, and the timestamp in ProbeRetry is less than or equal to the current time, a Probe option should be sent in the datagram, and the ProbeRetry field incremented by ProbeRetryTime.

Whether or not the Probe MTU option is sent in a datagram, if the ProbeReply bit is set, then a Reply MTU option with the value of the ReplyMTU field is placed in the outbound datagram. The ProbeReply bit is then cleared.

Every time a Probe option is sent, the ConsecutiveProbes value should be incremented. If this value reaches Max_ConsecutiveProbes, the SupportsProbe bit should be cleared.

When an IP datagram containing the Probe MTU option is received, the receiving IP sets the ReplyMTU to the Probe MTU option value and sets the ProbeReply bit in its outbound route to the source of the datagram. The SupportsProbe bit is set, and the ConsecutiveProbes value is reset to 0.

If an IP datagram containing the Reply MTU option is received, the IP layer must locate the routing cache entry corresponding to the source of the Reply MTU option; if no such entry exists, a new one (with default values) should be created. The SupportsProbe bit is set, and the ConsecutiveProbes value is reset to 0. The ProbeRetry field is set to the current time plus ReprobeInterval.

Four cases are possible when a Reply MTU option is received:

(1) The Reply MTU option value is less than the current

MINMTU: the MINMTU field is set to the new value, and

the LastDecreased field is set to the current time.

(2) The Reply MTU option value is greater than the

current MINMTU and the LastDecreased field plus IncreaseInterval is less than the current time: set the ProbeRetry field to LastDecreased plus IncreaseInterval, but do not change MINMTU.

(3) The Reply MTU option value is greater than the

current MINMTU and the LastDecreased field plus

IncreaseInterval is greater than the current time: set

the MINMTU field to the new value.

(4) The Reply MTU option value is equal to the current

MINMTU:

do nothing more.

Whenever the MTU field is changed, the transport layer should be notified, either by an upcall or by a change in a shared variable (which may be accessed from the transport layer by a downcall).

If a fragmentation reassembly timeout occurs, if an ICMP Time Exceeded/Fragmentation Reassembly Timeout is received, or if the IP layer is asked to send a probe by a higher layer, the ProbeRetry field for the appropriate routing cache entry is set to the current time. This will cause a Probe option to be sent with the next datagram (unless the SupportsProbe bit is turned off).

MANAGEMENT PARAMETERS

We suggest that the following parameters be made available to local applications and remote network management systems:

(1) The number of probe retries to be made before determining

a system is down. The value of 10 is certain to be wrong in some situations.

(2) The frequency with which probes are sent. Systems may

find that more or less frequent probing is more cost effective.

(3) The default MINMTU used to initialize routes.

(4) Applications should have the ability to force a probe

on a particular route. There are cases where a probe needs to be sent but the sender doesn't know it. An operator must be able to cause a probe in such situations. Furthermore, it may be useful for applications to "ping" for the MTU.

REFERENCES

   [1]  Kent, C. and J. Mogul, "Fragmentation Considered
        Harmful", Proc. ACM SIGCOMM '87, Stowe, VT, August 1987.
   
   [2]  Postel, J., Ed., "Internet Protocol", RFC-791,
        USC/Information Sciences Institute, Marina del Rey, CA,
        September 1981.
   
   [3]  Postel, J., Ed., "Transmission Control Protocol", RFC-793,
        USC/Information Sciences Institute, Marina del Rey, CA,
        September 1981.
   
   [4]  Postel, J., "The TCP Maximum Segment Size and Related Topics",
        RFC-879, USC/Information Sciences Institute, Marina del Rey,
        CA, November 1983.