Network Working Group J. Mogul Request for Comments: 1191 DECWRL Obsoletes: RFC 1063 S. Deering Stanford University November 1990
Path MTU Discovery
Status of this Memo
This RFC specifies a protocol on the IAB Standards Track for the Internet community, and requests discussion and suggestions for improvements. Please refer to the current edition of the "IAB Official Protocol Standards" for the standardization state and status of this protocol. Distribution of this memo is unlimited.
Table of Contents
Status of this Memo 1 Abstract 2 Acknowledgements 2 1. Introduction 2 2. Protocol overview 3 3. Host specification 4 3.1. TCP MSS Option 5 4. Router specification 6 5. Host processing of old-style messages 7 6. Host implementation 8 6.1. Layering 9 6.2. Storing PMTU information 10 6.3. Purging stale PMTU information 11 6.4. TCP layer actions 13 6.5. Issues for other transport protocols 14 6.6. Management interface 15 7. Likely values for Path MTUs 15 7.1. A better way to detect PMTU increases 16 8. Security considerations 18 References 18 Authors' Addresses 19
This memo describes a technique for dynamically discovering the maximum transmission unit (MTU) of an arbitrary internet path. It specifies a small change to the way routers generate one type of ICMP message. For a path that passes through a router that has not been so changed, this technique might not discover the correct Path MTU, but it will always choose a Path MTU as accurate as, and in many cases more accurate than, the Path MTU that would be chosen by current practice.
This proposal is a product of the IETF MTU Discovery Working Group.
The mechanism proposed here was first suggested by Geof Cooper , who in two short paragraphs set out all the basic ideas that took the Working Group months to reinvent.
When one IP host has a large amount of data to send to another host, the data is transmitted as a series of IP datagrams. It is usually preferable that these datagrams be of the largest size that does not require fragmentation anywhere along the path from the source to the destination. (For the case against fragmentation, see .) This datagram size is referred to as the Path MTU (PMTU), and it is equal to the minimum of the MTUs of each hop in the path. A shortcoming of the current Internet protocol suite is the lack of a standard mechanism for a host to discover the PMTU of an arbitrary path.
Note: The Path MTU is what in  is called the "Effective MTU for sending" (EMTU_S). A PMTU is associated with a path, which is a particular combination of IP source and destination address and perhaps a Type-of-service (TOS).
The current practice  is to use the lesser of 576 and the first-hop MTU as the PMTU for any destination that is not connected to the same network or subnet as the source. In many cases, this results in the use of smaller datagrams than necessary, because many paths have a PMTU greater than 576. A host sending datagrams much smaller than the Path MTU allows is wasting Internet resources and probably getting suboptimal throughput. Furthermore, current practice does not prevent fragmentation in all cases, since there are some paths whose PMTU is less than 576.
It is expected that future routing protocols will be able to provide accurate PMTU information within a routing area, although perhaps not across multi-level routing hierarchies. It is not clear how soon that will be ubiquitously available, so for the next several years the Internet needs a simple mechanism that discovers PMTUs without wasting resources and that works before all hosts and routers are modified.
In this memo, we describe a technique for using the Don't Fragment (DF) bit in the IP header to dynamically discover the PMTU of a path. The basic idea is that a source host initially assumes that the PMTU of a path is the (known) MTU of its first hop, and sends all datagrams on that path with the DF bit set. If any of the datagrams are too large to be forwarded without fragmentation by some router along the path, that router will discard them and return ICMP Destination Unreachable messages with a code meaning "fragmentation needed and DF set" . Upon receipt of such a message (henceforth called a "Datagram Too Big" message), the source host reduces its assumed PMTU for the path.
The PMTU discovery process ends when the host's estimate of the PMTU is low enough that its datagrams can be delivered without fragmentation. Or, the host may elect to end the discovery process by ceasing to set the DF bit in the datagram headers; it may do so, for example, because it is willing to have datagrams fragmented in some circumstances. Normally, the host continues to set DF in all datagrams, so that if the route changes and the new PMTU is lower, it will be discovered.
Unfortunately, the Datagram Too Big message, as currently specified, does not report the MTU of the hop for which the rejected datagram was too big, so the source host cannot tell exactly how much to reduce its assumed PMTU. To remedy this, we propose that a currently unused header field in the Datagram Too Big message be used to report the MTU of the constricting hop. This is the only change specified for routers in support of PMTU Discovery.
The PMTU of a path may change over time, due to changes in the routing topology. Reductions of the PMTU are detected by Datagram Too Big messages, except on paths for which the host has stopped setting the DF bit. To detect increases in a path's PMTU, a host periodically increases its assumed PMTU (and if it had stopped, resumes setting the DF bit). This will almost always result in datagrams being discarded and Datagram Too Big messages being
generated, because in most cases the PMTU of the path will not have changed, so it should be done infrequently.
Since this mechanism essentially guarantees that host will not receive any fragments from a peer doing PMTU Discovery, it may aid in interoperating with certain hosts that (improperly) are unable to reassemble fragmented datagrams.
When a host receives a Datagram Too Big message, it MUST reduce its estimate of the PMTU for the relevant path, based on the value of the Next-Hop MTU field in the message (see section 4). We do not specify the precise behavior of a host in this circumstance, since different applications may have different requirements, and since different implementation architectures may favor different strategies.
We do require that after receiving a Datagram Too Big message, a host MUST attempt to avoid eliciting more such messages in the near future. The host may either reduce the size of the datagrams it is sending along the path, or cease setting the Don't Fragment bit in the headers of those datagrams. Clearly, the former strategy may continue to elicit Datagram Too Big messages for a while, but since each of these messages (and the dropped datagrams they respond to) consume Internet resources, the host MUST force the PMTU Discovery process to converge.
Hosts using PMTU Discovery MUST detect decreases in Path MTU as fast as possible. Hosts MAY detect increases in Path MTU, but because doing so requires sending datagrams larger than the current estimated PMTU, and because the likelihood is that the PMTU will not have increased, this MUST be done at infrequent intervals. An attempt to detect an increase (by sending a datagram larger than the current estimate) MUST NOT be done less than 5 minutes after a Datagram Too Big message has been received for the given destination, or less than 1 minute after a previous, successful attempted increase. We recommend setting these timers at twice their minimum values (10 minutes and 2 minutes, respectively).
Hosts MUST be able to deal with Datagram Too Big messages that do not include the next-hop MTU, since it is not feasible to upgrade all the routers in the Internet in any finite time. A Datagram Too Big message from an unmodified router can be recognized by the presence of a zero in the (newly-defined) Next-Hop MTU field. (This is required by the ICMP specification , which says that "unused" fields must be zero.) In section 5, we discuss possible strategies
for a host to follow in response to an old-style Datagram Too Big message (one sent by an unmodified router).
A host MUST never reduce its estimate of the Path MTU below 68 octets.
A host MUST not increase its estimate of the Path MTU in response to the contents of a Datagram Too Big message. A message purporting to announce an increase in the Path MTU might be a stale datagram that has been floating around in the Internet, a false packet injected as part of a denial-of-service attack, or the result of having multiple paths to the destination.
A host doing PMTU Discovery must obey the rule that it not send IP datagrams larger than 576 octets unless it has permission from the receiver. For TCP connections, this means that a host must not send datagrams larger than 40 octets plus the Maximum Segment Size (MSS) sent by its peer.
Note: The TCP MSS is defined to be the relevant IP datagram size minus 40 . The default of 576 octets for the maximum IP datagram size yields a default of 536 octets for the TCP MSS.
Section 22.214.171.124 of "Requirements for Internet Hosts -- Communication Layers"  says:
Some TCP implementations send an MSS option only if the destination host is on a non-connected network. However, in general the TCP layer may not have the appropriate information to make this decision, so it is preferable to leave to the IP layer the task of determining a suitable MTU for the Internet path.
Actually, many TCP implementations always send an MSS option, but set the value to 536 if the destination is non-local. This behavior was correct when the Internet was full of hosts that did not follow the rule that datagrams larger than 576 octets should not be sent to non-local destinations. Now that most hosts do follow this rule, it is unnecessary to limit the value in the TCP MSS option to 536 for non-local peers.
Moreover, doing this prevents PMTU Discovery from discovering PMTUs larger than 576, so hosts SHOULD no longer lower the value they send
in the MSS option. The MSS option should be 40 octets less than the size of the largest datagram the host is able to reassemble (MMS_R, as defined in ); in many cases, this will be the architectural limit of 65495 (65535 - 40) octets. A host MAY send an MSS value derived from the MTU of its connected network (the maximum MTU over its connected networks, for a multi-homed host); this should not cause problems for PMTU Discovery, and may dissuade a broken peer from sending enormous datagrams.
Note: At the moment, we see no reason to send an MSS greater than the maximum MTU of the connected networks, and we recommend that hosts do not use 65495. It is quite possible that some IP implementations have sign-bit bugs that would be tickled by unnecessary use of such a large MSS.
When a router is unable to forward a datagram because it exceeds the MTU of the next-hop network and its Don't Fragment bit is set, the router is required to return an ICMP Destination Unreachable message to the source of the datagram, with the Code indicating "fragmentation needed and DF set". To support the Path MTU Discovery technique specified in this memo, the router MUST include the MTU of that next-hop network in the low-order 16 bits of the ICMP header field that is labelled "unused" in the ICMP specification . The high-order 16 bits remain unused, and MUST be set to zero. Thus, the message has the following format:
The size in octets of the largest datagram that could be forwarded, along the path of the original datagram, without being fragmented at this router. The size includes the IP header and IP data, and does not include any lower-level headers.
In this section we outline several possible strategies for a host to follow upon receiving a Datagram Too Big message from an unmodified router (i.e., one where the Next-Hop MTU field is zero). This section is not part of the protocol specification.
The simplest thing for a host to do in response to such a message is to assume that the PMTU is the minimum of its currently-assumed PMTU and 576, and to stop setting the DF bit in datagrams sent on that path. Thus, the host falls back to the same PMTU as it would choose under current practice (see section 3.3.3 of "Requirements for Internet Hosts -- Communication Layers" ). This strategy has the advantage that it terminates quickly, and does no worse than existing practice. It fails, however, to avoid fragmentation in some cases, and to make the most efficient utilization of the internetwork in other cases.
More sophisticated strategies involve "searching" for an accurate PMTU estimate, by continuing to send datagrams with the DF bit while varying their sizes. A good search strategy is one that obtains an accurate estimate of the Path MTU without causing many packets to be lost in the process.
Several possible strategies apply algorithmic functions to the previous PMTU estimate to generate a new estimate. For example, one could multiply the old estimate by a constant (say, 0.75). We do NOT recommend this; it either converges far too slowly, or it substantially underestimates the true PMTU.
A more sophisticated approach is to do a binary search on the packet size. This converges somewhat faster, although it still takes 4 or 5 steps to converge from an FDDI MTU to an Ethernet MTU. A serious disadvantage is that it requires a complex implementation in order to recognize when a datagram has made it to the other end (indicating that the current estimate is too low). We also do not recommend this strategy.
One strategy that appears to work quite well starts from the observation that there are, in practice, relatively few MTU values in use in the Internet. Thus, rather than blindly searching through arbitrarily chosen values, we can search only the ones that are
likely to appear. Moreover, since designers tend to chose MTUs in similar ways, it is possible to collect groups of similar MTU values and use the lowest value in the group as our search "plateau". (It is clearly better to underestimate an MTU by a few per cent than to overestimate it by one octet.)
In section 7, we describe how we arrived at a table of representative MTU plateaus for use in PMTU estimation. With this table, convergence is as good as binary search in the worst case, and is far better in common cases (for example, it takes only two round-trip times to go from an FDDI MTU to an Ethernet MTU). Since the plateaus lie near powers of two, if an MTU is not represented in this table, the algorithm will not underestimate it by more than a factor of 2.
Any search strategy must have some "memory" of previous estimates in order to chose the next one. One approach is to use the currently-cached estimate of the Path MTU, but in fact there is better information available in the Datagram Too Big message itself. All ICMP Destination Unreachable messages, including this one, contain the IP header of the original datagram, which contains the Total Length of the datagram that was too big to be forwarded without fragmentation. Since this Total Length may be less than the current PMTU estimate, but is nonetheless larger than the actual PMTU, it may be a good input to the method for choosing the next PMTU estimate.
Note: routers based on implementations derived from 4.2BSD Unix send an incorrect value for the Total Length of the original IP datagram. The value sent by these routers is the sum of the original Total Length and the original Header Length (expressed in octets). Since it is impossible for the host receiving such a Datagram Too Big message to know if it sent by one of these routers, the host must be conservative and assume that it is. If the Total Length field returned is not less than the current PMTU estimate, it must be reduced by 4 times the value of the returned Header Length field.
The strategy we recommend, then, is to use as the next PMTU estimate the greatest plateau value that is less than the returned Total Length field (corrected, if necessary, according to the Note above).
In the IP architecture, the choice of what size datagram to send is made by a protocol at a layer above IP. We refer to such a protocol as a "packetization protocol". Packetization protocols are usually transport protocols (for example, TCP) but can also be higher-layer protocols (for example, protocols built on top of UDP).
Implementing PMTU Discovery in the packetization layers simplifies some of the inter-layer issues, but has several drawbacks: the implementation may have to be redone for each packetization protocol, it becomes hard to share PMTU information between different packetization layers, and the connection-oriented state maintained by some packetization layers may not easily extend to save PMTU information for long periods.
We therefore believe that the IP layer should store PMTU information and that the ICMP layer should process received Datagram Too Big messages. The packetization layers must still be able to respond to changes in the Path MTU, by changing the size of the datagrams they send, and must also be able to specify that datagrams are sent with the DF bit set. We do not want the IP layer to simply set the DF bit in every packet, since it is possible that a packetization layer, perhaps a UDP application outside the kernel, is unable to change its datagram size. Protocols involving intentional fragmentation, while inelegant, are sometimes successful (NFS being the primary example), and we do not want to break such protocols.
To support this layering, packetization layers require an extension of the IP service interface defined in :
A way to learn of changes in the value of MMS_S, the "maximum send transport-message size", which is derived from the Path MTU by subtracting the minimum IP header size.
In general, the IP layer should associate each PMTU value that it has learned with a specific path. A path is identified by a source address, a destination address and an IP type-of-service. (Some implementations do not record the source address of paths; this is acceptable for single-homed hosts, which have only one possible source address.)
Note: Some paths may be further distinguished by different security classifications. The details of such classifications are beyond the scope of this memo.
The obvious place to store this association is as a field in the routing table entries. A host will not have a route for every possible destination, but it should be able to cache a per-host route for every active destination. (This requirement is already imposed by the need to process ICMP Redirect messages.)
When the first packet is sent to a host for which no per-host route exists, a route is chosen either from the set of per-network routes, or from the set of default routes. The PMTU fields in these route entries should be initialized to be the MTU of the associated first-hop data link, and must never be changed by the PMTU Discovery process. (PMTU Discovery only creates or changes entries for per-host routes). Until a Datagram Too Big message is received, the PMTU associated with the initially-chosen route is presumed to be accurate.
When a Datagram Too Big message is received, the ICMP layer determines a new estimate for the Path MTU (either from a non-zero Next-Hop MTU value in the packet, or using the method described in section 5). If a per-host route for this path does not exist, then one is created (almost as if a per-host ICMP Redirect is being processed; the new route uses the same first-hop router as the current route). If the PMTU estimate associated with the per-host route is higher than the new estimate, then the value in the routing entry is changed.
The packetization layers must be notified about decreases in the PMTU. Any packetization layer instance (for example, a TCP connection) that is actively using the path must be notified if the PMTU estimate is decreased.
Note: even if the Datagram Too Big message contains an Original Datagram Header that refers to a UDP packet, the TCP layer must be notified if any of its connections use the given
Also, the instance that sent the datagram that elicited the Datagram Too Big message should be notified that its datagram has been dropped, even if the PMTU estimate has not changed, so that it may retransmit the dropped datagram.
Note: The notification mechanism can be analogous to the mechanism used to provide notification of an ICMP Source Quench message. In some implementations (such as 4.2BSD-derived systems), the existing notification mechanism is not able to identify the specific connection involved, and so an additional mechanism is necessary.
Alternatively, an implementation can avoid the use of an asynchronous notification mechanism for PMTU decreases by postponing notification until the next attempt to send a datagram larger than the PMTU estimate. In this approach, when an attempt is made to SEND a datagram with the DF bit set, and the datagram is larger than the PMTU estimate, the SEND function should fail and return a suitable error indication. This approach may be more suitable to a connectionless packetization layer (such as one using UDP), which (in some implementations) may be hard to "notify" from the ICMP layer. In this case, the normal timeout-based retransmission mechanisms would be used to recover from the dropped datagrams.
It is important to understand that the notification of the packetization layer instances using the path about the change in the PMTU is distinct from the notification of a specific instance that a packet has been dropped. The latter should be done as soon as practical (i.e., asynchronously from the point of view of the packetization layer instance), while the former may be delayed until a packetization layer instance wants to create a packet. Retransmission should be done for only for those packets that are known to be dropped, as indicated by a Datagram Too Big message.
Internetwork topology is dynamic; routes change over time. The PMTU discovered for a given destination may be wrong if a new route comes into use. Thus, PMTU information cached by a host can become stale.
Because a host using PMTU Discovery always sets the DF bit, if the stale PMTU value is too large, this will be discovered almost
immediately once a datagram is sent to the given destination. No such mechanism exists for realizing that a stale PMTU value is too small, so an implementation should "age" cached values. When a PMTU value has not been decreased for a while (on the order of 10 minutes), the PMTU estimate should be set to the first-hop data-link MTU, and the packetization layers should be notified of the change. This will cause the complete PMTU Discovery process to take place again.
Note: an implementation should provide a means for changing the timeout duration, including setting it to "infinity". For example, hosts attached to an FDDI network which is then attached to the rest of the Internet via a slow serial line are never going to discover a new non-local PMTU, so they should not have to put up with dropped datagrams every 10 minutes.
An upper layer MUST not retransmit datagrams in response to an increase in the PMTU estimate, since this increase never comes in response to an indication of a dropped datagram.
One approach to implementing PMTU aging is to add a timestamp field to the routing table entry. This field is initialized to a "reserved" value, indicating that the PMTU has never been changed. Whenever the PMTU is decreased in response to a Datagram Too Big message, the timestamp is set to the current time.
Once a minute, a timer-driven procedure runs through the routing table, and for each entry whose timestamp is not "reserved" and is older than the timeout interval:
- The PMTU estimate is set to the MTU of the associated first hop.
- Packetization layers using this route are notified of the increase.
PMTU estimates may disappear from the routing table if the per-host routes are removed; this can happen in response to an ICMP Redirect message, or because certain routing-table daemons delete old routes after several minutes. Also, on a multi-homed host a topology change may result in the use of a different source interface. When this happens, if the packetization layer is not notified then it may continue to use a cached PMTU value that is now too small. One solution is to notify the packetization layer of a possible PMTU change whenever a Redirect message causes a route change, and whenever a route is simply deleted from the routing table.
The TCP layer must track the PMTU for the destination of a connection; it should not send datagrams that would be larger than this. A simple implementation could ask the IP layer for this value (using the GET_MAXSIZES interface described in ) each time it created a new segment, but this could be inefficient. Moreover, TCP implementations that follow the "slow-start" congestion-avoidance algorithm  typically calculate and cache several other values derived from the PMTU. It may be simpler to receive asynchronous notification when the PMTU changes, so that these variables may be updated.
A TCP implementation must also store the MSS value received from its peer (which defaults to 536), and not send any segment larger than this MSS, regardless of the PMTU. In 4.xBSD-derived implementations, this requires adding an additional field to the TCP state record.
Finally, when a Datagram Too Big message is received, it implies that a datagram was dropped by the router that sent the ICMP message. It is sufficient to treat this as any other dropped segment, and wait until the retransmission timer expires to cause retransmission of the segment. If the PMTU Discovery process requires several steps to estimate the right PMTU, this could delay the connection by many round-trip times.
Alternatively, the retransmission could be done in immediate response to a notification that the Path MTU has changed, but only for the specific connection specified by the Datagram Too Big message. The datagram size used in the retransmission should, of course, be no larger than the new PMTU.
Note: One MUST not retransmit in response to every Datagram Too Big message, since a burst of several oversized segments will give rise to several such messages and hence several retransmissions of the same data. If the new estimated PMTU is still wrong, the process repeats, and there is an exponential growth in the number of superfluous segments sent!
This means that the TCP layer must be able to recognize when a Datagram Too Big notification actually decreases the PMTU that it has already used to send a datagram on the given connection, and should ignore any other notifications.
Modern TCP implementations incorporate "congestion advoidance" and "slow-start" algorithms to improve performance . Unlike a retransmission caused by a TCP retransmission timeout, a retransmission caused by a Datagram Too Big message should not change the congestion window. It should, however, trigger the slow-start mechanism (i.e., only one segment should be retransmitted until acknowledgements begin to arrive again).
TCP performance can be reduced if the sender's maximum window size is not an exact multiple of the segment size in use (this is not the congestion window size, which is always a multiple of the segment size). In many system (such as those derived from 4.2BSD), the segment size is often set to 1024 octets, and the maximum window size (the "send space") is usually a multiple of 1024 octets, so the proper relationship holds by default. If PMTU Discovery is used, however, the segment size may not be a submultiple of the send space, and it may change during a connection; this means that the TCP layer may need to change the transmission window size when PMTU Discovery changes the PMTU value. The maximum window size should be set to the greatest multiple of the segment size (PMTU - 40) that is less than or equal to the sender's buffer space size.
PMTU Discovery does not affect the value sent in the TCP MSS option, because that value is used by the other end of the connection, which may be using an unrelated PMTU value.
Some transport protocols (such as ISO TP4 ) are not allowed to repacketize when doing a retransmission. That is, once an attempt is made to transmit a datagram of a certain size, its contents cannot be split into smaller datagrams for retransmission. In such a case, the original datagram should be retransmitted without the DF bit set, allowing it to be fragmented as necessary to reach its destination. Subsequent datagrams, when transmitted for the first time, should be no larger than allowed by the Path MTU, and should have the DF bit set.
The Sun Network File System (NFS) uses a Remote Procedure Call (RPC) protocol  that, in many cases, sends datagrams that must be fragmented even for the first-hop link. This might improve performance in certain cases, but it is known to cause reliability and performance problems, especially when the client and server are separated by routers.
We recommend that NFS implementations use PMTU Discovery whenever
routers are involved. Most NFS implementations allow the RPC datagram size to be changed at mount-time (indirectly, by changing the effective file system block size), but might require some modification to support changes later on.
Also, since a single NFS operation cannot be split across several UDP datagrams, certain operations (primarily, those operating on file names and directories) require a minimum datagram size that may be larger than the PMTU. NFS implementations should not reduce the datagram size below this threshold, even if PMTU Discovery suggests a lower value. (Of course, in this case datagrams should not be sent with DF set.)
We suggest that an implementation provide a way for a system utility program to:
- Specify that PMTU Discovery not be done on a given route.
- Change the PMTU value associated with a given route.
The former can be accomplished by associating a flag with the routing entry; when a packet is sent via a route with this flag set, the IP layer leaves the DF bit clear no matter what the upper layer requests.
These features might be used to work around an anomalous situation, or by a routing protocol implementation that is able to obtain Path MTU values.
The implementation should also provide a way to change the timeout period for aging stale PMTU information.
The algorithm recommended in section 5 for "searching" the space of Path MTUs is based on a table of values that severely restricts the search space. We describe here a table of MTU values that, as of this writing, represents all major data-link technologies in use in the Internet.
In table 7-1, data links are listed in order of decreasing MTU, and grouped so that each set of similar MTUs is associated with a "plateau" equal to the lowest MTU in the group. (The table also
includes some entries not currently associated with a data link, and gives references where available). Where a plateau represents more than one MTU, the table shows the maximum inaccuracy associated with the plateau, as a percentage.
We do not expect that the values in the table, especially for higher MTU levels, are going to be valid forever. The values given here are an implementation suggestion, NOT a specification or requirement. Implementors should use up-to-date references to pick a set of plateaus; it is important that the table not contain too many entries or the process of searching for a PMTU might waste Internet resources. Implementors should also make it convenient for customers without source code to update the table values in their systems (for example, the table in a BSD-derived Unix kernel could be changed using a new "ioctl" command).
Note: It might be a good idea to add a few table entries for values equal to small powers of 2 plus 40 (for the IP and TCP headers), where no similar values exist, since this seems to be a reasonably non-arbitrary way of choosing arbitrary values.
The table might also contain entries for values slightly less than large powers of 2, in case MTUs are defined near those values (it is better in this case for the table entries to be low than to be high, or else the next lowest plateau may be chosen instead).
Section 6.3 suggests detecting increases in the PMTU value by periodically increasing the PTMU estimate to the first-hop MTU. Since it is likely that this process will simply "rediscover" the current PTMU estimate, at the cost of several dropped datagrams, it should not be done often.
A better approach is to periodically increase the PMTU estimate to the next-highest value in the plateau table (or the first-hop MTU, if that is smaller). If the increased estimate is wrong, at most one round-trip time is wasted before the correct value is rediscovered. If the increased estimate is still too low, a higher estimate will be attempted somewhat later.
Because it may take several such periods to discover a significant increase in the PMTU, we recommend that a short timeout period should be used after the estimate is increased, and a longer timeout be used
after the PTMU estimate is decreased because of a Datagram Too Big message. For example, after the PTMU estimate is decreased, the timeout should be set to 10 minutes; once this timer expires and a larger MTU is attempted, the timeout can be set to a much smaller value (say, 2 minutes). In no case should the timeout be shorter than the estimated round-trip time, if this is known.
This Path MTU Discovery mechanism makes possible two denial-of- service attacks, both based on a malicious party sending false Datagram Too Big messages to an Internet host.
In the first attack, the false message indicates a PMTU much smaller than reality. This should not entirely stop data flow, since the victim host should never set its PMTU estimate below the absolute minimum, but at 8 octets of IP data per datagram, progress could be slow.
In the other attack, the false message indicates a PMTU greater than reality. If believed, this could cause temporary blockage as the victim sends datagrams that will be dropped by some router. Within one round-trip time, the host would discover its mistake (receiving Datagram Too Big messages from that router), but frequent repetition of this attack could cause lots of datagrams to be dropped. A host, however, should never raise its estimate of the PMTU based on a Datagram Too Big message, so should not be vulnerable to this attack.
A malicious party could also cause problems if it could stop a victim from receiving legitimate Datagram Too Big messages, but in this case there are simpler denial-of-service attacks available.
 R. Braden, ed. Requirements for Internet Hosts -- Communication Layers. RFC 1122, SRI Network Information Center, October, 1989.
 Geof Cooper. IP Datagram Sizes. Electronic distribution of the TCP-IP Discussion Group, Message-ID <8705240517.AA01407@apolling.imagen.uucp>.
 ISO. ISO Transport Protocol Specification: ISO DP 8073. RFC 905, SRI Network Information Center, April, 1984.
 Van Jacobson. Congestion Avoidance and Control. In Proc. SIGCOMM '88 Symposium on Communications Architectures and Protocols, pages 314-329. Stanford, CA, August, 1988.
 C. Kent and J. Mogul. Fragmentation Considered Harmful. In Proc. SIGCOMM '87 Workshop on Frontiers in Computer Communications Technology. August, 1987.