Boeing Research & Technology

P.O. Box 3707 Seattle WA 98124 USA fltemplin@acm.org

I-D Internet-Draft IP packets (both IPv4 and IPv6) contain a single unit of upper layer protocol data which becomes the retransmission unit in case of loss. Upper layer protocols including the Transmission Control Protocol (TCP) and transports over the User Datagram Protocol (UDP) prepare data units known as "segments", with traditional arrangements including a single segment per IP packet. This document presents a new construct known as the "IP Parcel" which permits a single packet to carry multiple upper layer protocol segments, essentially creating a "packet-of-packets". IP parcels provide an essential building block for improved performance and efficiency by supporting larger Maximum Transmission Units (MTUs) in the Internet as discussed in this document.

IP packets (both IPv4 and IPv6 ) contain a single unit of upper layer protocol data which becomes the retransmission unit in case of loss. Upper layer protocols such as the Transmission Control Protocol (TCP) and transports over the User Datagram Protocol (UDP) (including QUIC , LTP and others) prepare data units known as "segments", with traditional arrangements including a single segment per IP packet. This document presents a new construct known as the "IP Parcel" which permits a single packet to carry multiple upper layer protocol segments. This essentially creates a "packet-of-packets" with the IP layer and full {TCP,UDP} headers appearing only once but with possibly multiple segments included. Parcels are formed when an upper layer protocol entity identified by the "5-tuple" (source address, destination address, source port, destination port, protocol number) prepares a data buffer beginning with a checksum header block followed by the concatenation of up to 64 properly-formed segments that can be broken out into smaller sub-parcels and/or individual packets using a copy of the {TCP,UDP}/IP headers if necessary. All segments except the final one must be equal in length and no larger than 65535 octets (minus headers), while the final segment must be no larger than the others but may be smaller. The upper layer protocol entity then delivers the buffer, number of segments and non-final segment size to lower layers which append a {TCP,UDP} header and an IP header plus extensions that identify this as a parcel and not an ordinary packet. Parcels can be forwarded over consecutive parcel-capable links in the path until arriving at a router with a next hop link that does not support parcels or an ingress middlebox OMNI interface that spans intermediate Internetworks using adaptation layer encapsulation and fragmentation. In the former case, the router transforms the parcel into individual IP packets or smaller sub-parcels then forwards each via the next hop link. In the latter case, the OMNI interface breaks the parcel out into smaller sub-parcels if necessary then encapsulates each (sub-)parcel in headers suitable for traversing the Internetworks while applying fragmentation if necessary. These sub-parcels may then be recombined into one or more larger parcels by an egress middlebox OMNI interface which either delivers them locally or forwards them over additional parcel-capable links on the path to the final destination. Reordering and even loss or damage of individual segments in the network is therefore possible, but what matters is that the number of parcels delivered to the final destination should be kept to a minimum for the sake of efficiency and that the loss or receipt of individual segments (and not parcel size) determines the retransmission unit. The following sections discuss rationale for creating and shipping parcels as well as the actual protocol constructs and procedures involved. IP parcels provide an essential building block for improved performance and efficiency while supporting larger Maximum Transmission Units (MTUs) in the Internet. It is further expected that the parcel concept will drive future innovation in applications, operating systems, network equipment and data links.

The Oxford Languages dictionary defines a "parcel" as "a thing or collection of things wrapped in paper in order to be carried or sent by mail". Indeed, there are many examples of parcel delivery services worldwide that provide an essential transit backbone for efficient business and consumer transactions. In this same spirit, an "IP parcel" is simply a collection of up to 64 upper layer protocol segments wrapped in an efficient package for transmission and delivery (i.e., a "packet-of-packets") while a "singleton IP parcel" is simply a parcel that contains a single segment. IP parcels are distinguished from ordinary packets through the special header constructions discussed in this document. The IP parcel construct is defined for both IPv4 and IPv6. Where the document refers to "IPv4 header length", it means the total length of the base IPv4 header plus all included options, i.e., as determined by consulting the Internet Header Length (IHL) field. Where the document refers to "IPv6 header length", however, it means only the length of the base IPv6 header (i.e., 40 octets), while the length of any extension headers is referred to separately as the "IPv6 extension header length". Finally, the term "IP header plus extensions" refers generically to an IPv4 header plus all included options or an IPv6 header plus all included extension headers. Where the document refers to "upper layer header length", it means the length of either the UDP header (8 octets) or the TCP header plus options (20 octets or more). It is important to note that only a single IP header and a single full {TCP,UDP} header appears in each parcel regardless of the number of segments included. This distinction often provides a significant savings in overhead made possible only by IP parcels. Where the document refers to checksum calculations, it means the standard Internet checksum unless otherwise specified. The same as for TCP , UDP and IPv4 , the standard Internet checksum is defined as (sic) "the 16-bit one's complement of the one's complement sum of all (pseudo-)headers plus data, padded with zero octets at the end (if necessary) to make a multiple of two octets". A notional Internet checksum algorithm can be found in , with the understanding that practical implementations require special attention to byte ordering "endianness" to ensure interoperability between diverse architectures. Where the document refers to "parcel path MTU", it means the maximum-sized IP parcel that can traverse the forward path to the destination as determined through parcel path qualification (see: ). Note that this size may be larger than the maximum-sized singleton IP packet that can traverse the same path, since intermediate nodes can break oversized IP parcels into smaller sub-parcels but cannot do so for singleton IP packets. Finally, the term "parcel-capable link" refers to any data link medium (physical or virtual) capable of transiting a {TCP,UDP}/IP packet that employs the parcel-specific constructions specified in this document. The link MUST be capable of forwarding parcels with at least one segment of maximum size, therefore each parcel-capable link MUST configure an MTU of at least 64KB and SHOULD configure a larger MTU. Currently, only the OMNI link satisfies these properties. The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 when, and only when, they appear in all capitals, as shown here.

Studies have shown that applications can improve their performance by sending and receiving larger packets due to reduced numbers of system calls and interrupts as well as larger atomic data copies between kernel and user space. Larger packets also result in reduced numbers of network device interrupts and better network utilization (e.g., due to header overhead reduction) in comparison with smaller packets. A first study involved performance enhancement of the QUIC protocol using the linux Generic Segment/Receive Offload (GSO/GRO) facility. GSO/GRO provides a robust (but non-standard) service very similar in nature to the IP parcel service described here, and its application has shown significant performance increases due to the increased transfer unit size between the operating system kernel and QUIC application. Unlike IP parcels, however, GSO/GRO perform fragmentation and reassembly at the transport layer with the transport segment size limited by the path MTU. A second study showed that GSO/GRO also improves performance for the Licklider Transmission Protocol (LTP) used for the Delay Tolerant Networking (DTN) Bundle Protocol for segments larger than the actual path MTU through the use of fragmentation. Historically, the NFS protocol also saw significant performance increases using larger (single-segment) UDP datagrams even when IP fragmentation is invoked, and LTP still follows this profile today. Moreover, LTP shows this (single-segment) performance increase profile extending to the largest possible segment size which suggests that additional performance gains are possible using (multi-segment) IP parcels that approach or even exceed 65535 octets. TCP also benefits from larger packet sizes and efforts have investigated TCP performance using jumbograms internally with changes to the linux GSO/GRO facilities . The idea is to use the jumbo payload option internally and to allow GSO/GRO to use buffer sizes larger than 65535 octets, but with the understanding that links that support jumbos natively are not yet widely available. Hence, IP parcels provides a packaging that can be considered in the near term under current deployment limitations. A limiting consideration for sending large packets is that they are often lost at links with smaller MTUs, and the resulting Packet Too Big (PTB) message may be lost somewhere in the path back to the original source. This "Path MTU black hole" condition can degrade performance unless robust path probing techniques are used, however the best case performance always occurs when no packets are lost due to size restrictions. These considerations therefore motivate a design where transport protocols should employ a maximum segment size no larger than 65535 octets (minus headers), while parcels that carry multiple segments may themselves be significantly larger. Then, even if the network needs to sub-divide the parcels into smaller sub-parcels to forward further toward the final destination, an important performance optimization for the original source, final destination and network path as a whole can be realized. An analogy: when a consumer orders 50 small items from a major online retailer, the retailer does not ship the order in 50 separate small boxes. Instead, the retailer packs as many of the small items as possible into one or a few larger boxes (i.e., parcels) then places the parcels on a semi-truck or airplane. The parcels may then pass through one or more regional distribution centers where they may be repackaged into different parcel configurations and forwarded further until they are finally delivered to the consumer. But most often, the consumer will only find one or a few parcels at their doorstep and not 50 separate small boxes. This flexible parcel delivery service greatly reduces shipping and handling cost for all including the retailer, regional distribution centers and finally the consumer.

IP parcels are formed by an upper layer protocol entity (identified by the 5-tuple described above) when it prepares a data buffer containing the concatenation of a block of up to 64 Checksums followed by their corresponding upper layer protocol segments (with each TCP non-first segment preceded by a 4-octet Sequence Number). All non-final segments MUST be equal in length while the final segment MUST NOT be larger and MAY be smaller. Each non-final segment MUST NOT be larger than 65535 octets minus the length of the UDP header or TCP header and its options, minus the length of the IPv4/IPv6 header and its options/extensions minus 2 octets for the per-segment Checksum. (Note that this also satisfies the case of ingress middlebox OMNI interfaces in the path that would regard the headers as upper layer protocol payload during IPv6 encapsulation/fragmentation.) The upper layer protocol entity then presents the buffer and non-final segment size to lower layers, noting that the buffer may be larger than 65535 octets if it includes sufficient segments of a large enough size to exceed that value. If the buffer plus headers would together be no larger than the parcel path MTU, the lower layers then append a single full {TCP,UDP} header (plus options) followed by a single IPv4/IPv6 header plus options/extensions. If the buffer would cause a single parcel to exceed the parcel path MTU, lower layers instead break the buffer up into multiple smaller buffers (each with an integral number of segments) and append separate {TCP,UDP}/IP headers for each. The IP layer then presents each parcel to a network interface attachment to either an ordinary parcel-capable link or an OMNI link that performs adaptation layer encapsulation and fragmentation (see: ). The IP layer includes a Jumbo Payload option in the IP header formed as shown in :

| +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Option Type | Opt Data Len | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ | Nsegs | Jumbo Payload Length | +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ |<------------------------ Option Data ------------------------>|]]>

For IPv4, the Jumbo Payload option format follows from except that the IP layer sets option type to '00001011' and option length to '00000110' noting that the length distinguishes this type from its obsoleted use as the "IPv4 Probe MTU" option . The IP layer also interprets the most significant option data octet as an "Nsegs" field that encodes a value J between 1 and 64 and interprets the following 3-octets as a "Jumbo Payload Length" field that encodes the length of the IPv4 header plus the length of the {TCP,UDP} header (plus options) plus the combined length of the checksum block plus all concatenated segments. The IP layer next sets the IPv4 header DF bit to 1, then sets the IPv4 header Total Length field to the length of the first segment only. Note that the IP layer can form true IPv4 jumbograms (as opposed to parcels) by instead setting the IPv4 header Total Length field to 0 and treating the entire 4 octets of the option data as the Jumbo Payload Length (see: ). For IPv6, the IP layer includes a Jumbo Payload option in an IPv6 Hop-by-Hop Options extension header formatted the same as for IPv4 above, but with option type set to '11000010' and option length set to '00000100'. The IP layer sets the option data "Nsegs" field to a 1-octet value J between 1 and 64 and sets the "Jumbo Payload Length" field to a 3-octet value that encodes the lengths of all IPv6 extension headers present plus the length of the {TCP,UDP} header (plus options) plus the combined length of the checksum block plus all concatenated segments. The IP layer next sets the IPv6 header Payload Length field to the length of the first segment only. Note that the IP layer can form true IPv6 jumbograms (as opposed to parcels) by instead setting the IPv6 header Payload Length field to 0 and treating the entire 4 octets of the option data as the Jumbo Payload Length (see: ). {TCP,UDP}/IP parcel formats are shown in :

where J is the total number of segments (between 1 and 64), L is the length of each non-final segment which MUST NOT be larger than 65535 octets (minus headers) and K is the length of the final segment which MUST NOT be larger than L. For both TCP and UDP, the {TCP,UDP} header is immediately followed by a block of J 2-octet Checksums which are then followed by J upper layer protocol segments. For TCP, the TCP header sequence number field encodes a 4-octet starting sequence number for the first segment only, while each non-first segment is preceded by its own 4-octet Sequence Number field. For this reason, the length of the first TCP segment is only (L-4) octets since the 4 octet TCP header sequence number field applies to that segment. The {Total, Payload} Length is then set to L if there are multiple segments or K if there is only a single segment. Next, the 1-octet Nsegs field is set to J and the 3-octet Jumbo Payload Length "M" is set to the length of the IP header plus extensions for IPv4 (or to the length of the extension headers only for IPv6), plus the length of the UDP header or TCP header plus options, plus the lengths of the Checksum block and all concatenated Segments that follow. The Nsegs value unambiguously determines the number of 2-octet fields present in the checksum block and jointly determines the number of parcel data segments in conjunction with the Jumbo Payload Length. Receivers therefore observe the following robustness considerations: if Nsegs is less than 1 or greater than 64, or if the Jumbo Payload Length indicates insufficient space for the full checksum block plus at least one octet of data, the receiver discards the parcel. if the data length following the checksum block is less than (((Nsegs - 1) * L) + 1) the receiver processes all initial fields of the checksum block along with their corresponding segments up to the end of the data and ignores any remaining checksums. if the data length following the checksum block is greater than (Nsegs * L) the receiver processes all checksums with their corresponding segments and ignores any remaining data beyond the end of the final segment. Note: per-segment checksums appear in a contiguous data block immediately following the {TCP,UDP}/IP headers instead of inline with the parcel segments to greatly increase the likelihood that they will appear in the contiguous head of a kernel receive buffer even if the parcel was subject to OMNI interface IPv6 fragmentation. This condition may not always hold if the IPv6 fragments also incur IPv4 encapsulation and fragmentation over paths that traverse slow IPv4 links with small MTUs. In that case, performance is bounded by the unavoidable slow link traversal and not the overhead for pulling the fragmented checksum block into the contiguous head of a kernel receive buffer.

A TCP Parcel is an IP Parcel that includes an IP header plus extensions with a Jumbo Payload option encoding the number of segments (Nsegs) and Jumbo Payload length up to 4MB. The IP header plus extensions is then followed by a 20-octet TCP header plus any additional TCP option octets. The TCP header is then followed by J consecutive 2-octet Checksum fields followed by J consecutive segments, where the final segment is K octets in length, each non-first segment is L octets in length and includes a 4-octet Sequence Number and the first segment is (L-4) octets in length and uses the sequence number found in the TCP header. The segment length "L" is encoded in the IP header {Total, Payload} Length field while the number of segments J is encoded in the Nsegs octet. The overall length of the parcel as well as final segment length are determined by the Jumbo Payload length "M" as discussed above. The source prepares TCP Parcels in a similar fashion as for TCP jumbograms . The source calculates the checksum of the TCP header plus IP pseudo-header only, but with the TCP header sequence number field temporarily set to 0 during the calculation since the true sequence number will be included as a pseudo header for the first segment. The source then writes the calculated value in the TCP header checksum field as-is (i.e., without converting calculated '0' values to 'ffff') and finally re-writes the actual sequence number back into the sequence number field. (Nodes that verify the header checksum first perform the same operation of temporarily setting the sequence number field to 0 and then resetting to the actual value after checksum verification.) The source then calculates the checksum of the first segment beginning with the sequence number found in the full TCP header as a 4-octet pseudo-header and extending over the (L-4) octet length of the segment. The source next calculates the checksum for each non-first segment independently over the L octet length of the segment beginning with the per-segment Sequence Number. As the source calculates each per-segment checksum for segments i=(1 thru J), it writes the value into the corresponding segment Checksum(i) field with calculated '0' values written as 'ffff'. See: for further discussion.

A UDP Parcel is an IP Parcel that includes an IP header plus extensions with a Jumbo Payload option encoding the number of segments (Nsegs) and Jumbo Payload length up to 4MB. The IP header plus extensions is then followed by an 8-octet UDP header followed by J consecutive 2-octet Checksum fields followed by J upper layer protocol segments. Each segment must begin with a transport-specific start delimiter (e.g., a segment identifier) included by the transport layer user of UDP. The length "L" of each non-final segment is encoded in the IP {Total, Payload} Length field while the number of segments J is encoded in the Nsegs octet. The overall length of the parcel as well as the final segment length are determined by the Jumbo Payload length "M" as discussed above. The source prepares UDP Parcels in a similar fashion as for UDP jumbograms and MUST set the UDP header length field to 0. The source then calculates the checksum of the UDP header plus IP pseudo-header and writes the calculated value in the UDP header checksum field as-is (i.e., without converting calculated '0' values to 'ffff'). The source then calculates a separate checksum for each segment independently over the length of the segment. As the source calculates each per-segment checksum for segments i=(1 thru J), it writes the value into the corresponding Checksum(i) field with calculated '0' values written as 'ffff'. See: for further discussion.

The IP layer of the source next presents each parcel to a network interface for transmission over a parcel-capable link. For ordinary IP interface attachments to parcel-capable links, the interface simply admits each parcel into the link the same as for any IP packet after which it may then be forwarded by any number of routers over additional consecutive parcel-capable links possibly even traversing the entire forward path to the final destination itself. If any router in the path does not recognize the parcel construct, it drops the parcel and may return an ICMP "Parameter Problem" message. If the router recognizes parcels but the next hop link in the path does not, or if the parcel would exceed the parcel path MTU, the router instead opens the parcel. The router then forwards each enclosed segment in singleton IP packets or in a set of smaller sub-parcels that each contain a subset of all segments. The router prepares each singleton IP packet or smaller sub-parcel for transmission to the next hop as follows. For transmission of singleton IP packets over links that do not support parcels, the router removes the Jumbo Payload option and the per segment Checksum and Sequence Number fields then sets IP {Total, Payload} length according to the standards . For TCP, the router then sets the TCP header Sequence Number field based on the starting sequence number for the segment according to and also clears the ACK flag in all but the first packet. For UDP, the router then sets the UDP length field according to . For both TCP and UDP, the router next calculates the checksum over the length of the packet according to the native {TCP,UDP} protocol specification, then writes the value in the {TCP,UDP} header checksum field and finally forwards the packet. For transmission of smaller sub-parcels over parcel capable links, the router breaks the original parcel into smaller groups of segments that would fit within the parcel path MTU by determining the number of segments of length L that can fit into each sub-parcel under the size constraints. For example, if the router determines that a sub-parcel can contain 3 segments of length L, it creates sub-parcels with the first containing segments 1-3, the second containing segments 4-6, etc., and with the final containing any remaining segments. The router then appends identical {TCP,UDP}/IP headers plus extensions to each sub-parcel while resetting L and M in each according to the above equations with Nsegs (J) set to 3 (and K = L) for each non-final sub-parcel and with Nsegs set to the remaining number of segments for the final sub-parcel. For TCP, the router then sets the TCP Sequence Number field to the value that appears in the first sub-parcel segment while omitting the first segment Sequence Number header (if present) and also clearing the ACK flag in all sub-parcels except the first. For both TCP and UDP, the router finally resets the {TCP,UDP} header checksum according to ordinary parcel formation procedures (see above) then forwards each (sub-)parcel over the outgoing parcel-capable link. If the outgoing network interface for the original parcel or sub-parcel is an OMNI interface , the OMNI Adaptation Layer (OAL) of this First Hop Segment (FHS) OAL source node then forwards the parcel to the next OAL hop which may be either an OAL intermediate node or a Last Hop Segment (LHS) OAL destination node. Note that OMNI interface upper layer protocol processing procedures are specified in detail the remainder of this section, while lower layer encapsulation and fragmentation procedures are specified in detail in . When the OAL source forwards a parcel (whether generated by a local application or generated by another node then forwarded over one or more parcel capable links), it first assigns a monotonically-incrementing (modulo 127) "Parcel ID" for adaptation layer processing. If necessary, the OAL source then subdivides the parcel into sub-parcels the same as discussed for the IP layer parcel subdivision procedures discussed above. The OAL source next selects a monotonically-incrementing Identification value for each sub-parcel then performs adaptation layer encapsulation and fragmentation and finally forwards them to the next OAL hop which forwards further toward the OAL destination as necessary. When the sub-parcels arrive at the OAL destination, the node can optionally retain them along with their Parcel ID and Identifications for a brief time to support re-combining with peer sub-parcels of the same original parcel identified by the adaptation layer 4-tuple consisting of the (source, destination, Identification, Parcel ID) fields. This re-combining entails the concatenation of segments included in sub-parcels with the same Parcel ID and with Identification values within 64 of one another to create a larger sub-parcel possibly even as large as the entire original parcel. Order of concatenation need not be strictly enforced, with the exception that the sub-parcel containing the final segment must occur as a final concatenation and not as an intermediate. The OAL destination then appends a common {TCP,UDP}/IP header plus extensions to each re-combined sub-parcel while resetting J, K, L and M in each according to the above equations. For TCP, if any sub-parcels have the ACK bit set the OAL destination also sets the ACK bit in the re-combined sub-parcel TCP header. The OAL destination then resets the {TCP,UDP}/IP header checksum for each re-combined sub-parcel. If the OAL destination is also the final destination, it then delivers the sub-parcels to the IP layer which processes them according to the 5-tuple information supplied by the original source. Otherwise, the OAL destination forwards each sub-parcel toward the final destination the same as for an ordinary IP packet as discussed above. Note: sub-dividing a larger parcel into two or more sub-parcels entails replication of the {TCP,UDP}/IP headers. For TCP, the process entails copying the full TCP/IP header from the original parcel while writing the sequence number of the first sub-parcel segment into the TCP Sequence Number field, clearing the ACK bit if necessary as discussed above and truncating the (new) first segment Sequence Number field. For UDP, the process entails copying the full UDP/IP header from the original parcel into each sub-parcel. For both TCP and UDP, the process finally includes recalculating and resetting Nsegs and Jumbo Payload Length then recalculating the {TCP,UDP} header checksum. Note that the per-segment Checksum values in the sub-parcel segments themselves are still valid and need not be recalculated. Note: re-combining two or more sub-parcels into a larger parcel entails a reverse process of the above in which the {TCP,UDP}/IP headers of non-first sub-parcels are discarded and their included segments concatenated following those of a first sub-parcel. For TCP, the process includes setting the ACK in the TCP header only if ACK was set in any of the original sub-parcels. For both TCP and UDP, the process finally includes recalculating and resetting Nsegs and Jumbo Payload Length then recalculating the {TCP,UDP} header checksum as discussed above. Note that the per-segment Checksum values in the combined parcel segments themselves are still valid and need not be recalculated. (This process should not be performed by the OAL destination if it would negatively impact performance, noting that it is always acceptable to forward individual sub-parcels without attempting to re-combine them and without delay.) Note: while the OAL destination and/or final destination could theoretically re-combine the sub-parcels of multiple different parcels with identical upper layer protocol 5-tuples and non-final segment lengths, this process could become complicated when the different parcels each have differing final segment lengths. Since this might interfere with any perceived performance advantage, the decision of whether and how to perform inter-parcel concatenation is an implementation matter. Note: sub-dividing of IP parcels occurs only at OMNI link ingress nodes while re-combining of IP parcels occurs only at OMNI link egress nodes. Therefore, intermediate OAL nodes do not participate in the sub-dividing or recombining processes. For TCP, the ACK bit must be managed as specified above to avoid confusing receivers with spurious duplicate ACK indications.

To determine whether parcels are supported over at least a leading portion of the forward path up to and including the final destination, the original source can send IP parcels that contain Jumbo Payload options formatted as "Parcel Probes". The purpose of the probe is to elicit a "Parcel Reply" and possibly also an ordinary upper layer protocol probe reply from the final destination. The former is used to establish the parcel path MTU, while the latter determines the (transport layer) maximum segment size. If the original source receives a positive Parcel Reply, it marks the path as "parcels supported" and ignores any ordinary ICMP and/or Packet Too Big (PTB) messages concerning the probe. If the original source instead receives a negative Parcel Reply or no reply, it marks the path as "parcels not supported" and may regard any ordinary ICMP and/or PTB messages concerning the probe (or its contents) as indications of a possible MTU restriction. The original source can therefore send Parcel Probes in parallel with sending real data as ordinary IP packets/parcels. The parcel probes will traverse parcel-capable links joined by routers on the forward path possibly extending all the way to the destination. If the original source receives a Parcel Reply, it can continue using IP parcels. Parcel Probes use the same Jumbo Payload option type used for ordinary parcels (see: ) but set a different option length and include a 4-octet "(Parcel) Path MTU" field into which conformant routers write the minimum link MTU observed in a similar fashion as described in . Parcel Probes include one or more upper layer protocol segments corresponding to the 5-tuple for the flow, which may also include {TCP,UDP} segment size probes used for packetization layer path MTU discovery . The original source sends Parcel Probes unidirectionally in the forward path toward the final destination to elicit a Parcel Reply, since it will often be the case that IP parcels are supported only in the forward path and not in the return path. Parcel Probes may be filtered in the forward path by any node that does not recognize IP parcels, but Parcel Replys must be packaged to avoid filtering since parcels may not be recognized along portions of the return path. For this reason, the Jumbo Payload options included in Parcel Probes are always packaged as IPv4 header options or IPv6 Hop-by-Hop options while Parcel Replys are returned as UDP/IP encapsulated ICMPv6 PTB messages with a "Parcel Reply" Code value (see: ). Original sources send Parcel Probes that include a Jumbo Payload option coded in an alternate format as shown in :

For IPv4, the original source includes the option as an IPv4 option with Type set to '00001011' the same as for an ordinary IPv4 parcel (see: ) but with Length set to '00010100' to distinguish this as a probe. The original source sets Nonce-1 to '11111111', sets Check to the same value that will appear in the TTL of the outgoing IPv4 header, sets PMTU to the MTU of the outgoing IPv4 interface and sets Nonce-2 to a 64-bit random number. The source next includes a {TCP,UDP} header followed by upper layer protocol segments in the same format as for an ordinary parcel. The source then sets {Nsegs, Jumbo Payload Length, IPv4 Total Length} and calculates the header and per-segment checksums the same as for an ordinary parcel. The source finally sends the Parcel Probe via the outbound IPv4 interface. According to , middleboxes (i.e., routers, security gateways, firewalls, etc.) that do not observe this specification SHOULD drop IP packets that contain option type '00001011' ("IPv4 Probe MTU") but some might instead either attempt to implement or ignore the option altogether. IPv4 middleboxes that observe this specification instead MUST process the option as a Parcel Probe as specified below. For IPv6, the original source includes the probe option as an IPv6 Hop-by-Hop option with Type set to '11000010' the same as for an ordinary IPv6 parcel (see: ) but with Length set to '00010010' to distinguish this as a probe. The original source sets Nonce-1 to '11111111', sets Check to the same value that will appear in the Hop Limit of the outgoing IPv6 header, sets PMTU to the MTU of the outgoing IPv6 interface and sets Nonce-2 to a 64-bit random number. The source next includes a {TCP,UDP} header followed by one or more upper layer protocol segments in the same format as for an ordinary parcel. The source then sets {Nsegs, Jumbo Payload Length, IPv6 Payload Length} and calculates the header and per-segment checksums the same as for an ordinary parcel, then finally sends the Parcel Probe via the outbound IPv6 interface. According to , middleboxes (i.e., routers, security gateways, firewalls, etc.) that recognize the IPv6 Jumbo Payload option but do not observe this specification SHOULD return an ICMPv6 Parameter Problem message (and presumably also drop the packet) due to the different option length. IPv6 middleboxes that observe this specification instead MUST process the option as a Parcel Probe as specified below. When a router that observes this specification receives either an IPv4 or IPv6 Parcel Probe it first compares Nonce-1 with '11111111' and Check with the IP header TTL/Hop Limit; if either value differs, the router MUST drop the probe and return a negative Parcel Reply (see below). Otherwise, if the next hop link is non-parcel-capable or configures an MTU that is too small to pass the probe, the router compares the PMTU value with the MTU of the inbound link for the probe and MUST (re)set PMTU to the lower MTU. The router then MUST return a positive Parcel Reply (see below) and convert the probe into an ordinary IP packet(s) the same as was described previously for routers forwarding to non-parcel-capable links. If the next hop IP link configures a sufficiently large MTU to pass the packet(s), the router then MUST forward each packet to the next hop; otherwise, it MUST drop each packet and return a suitable PTB. If the next hop IP link both supports parcels and configures an MTU that is large enough to pass the probe, the router instead compares the probe PMTU value with the MTUs of both the inbound and outbound links for the probe and MUST (re)set PMTU to the lower MTU. The router then MUST reset Check to the same value that will appear in the TTL/Hop Limit of the outgoing IP header, and MUST forward the Parcel Probe to the next hop. The final destination may therefore receive either one or more ordinary IP packets or an intact Parcel Probe. If the final destination receives ordinary IP packets, it performs any necessary integrity checks then delivers the packets to upper layers which will return an upper layer probe response if necessary. If the final destination receives a Parcel Probe, it first compares Nonce-1 with '11111111' and Check with the IP header TTL/Hop Limit; if either value differs, the final destination MUST drop the probe and return a negative Parcel Reply. Otherwise, the final destination compares the probe PMTU value with the MTU of the inbound link and MUST (re)set PMTU to the lower MTU. The final destination then MUST return a positive Parcel Reply and deliver the probe contents to upper layers the same as for an ordinary IP parcel. When a router or final destination returns a Parcel Reply, it prepares an ICMPv6 PTB message with Code set to "Parcel Reply" and with MTU set to either the PMTU value reported in the Parcel Probe for a positive reply or to the value 0 for a negative reply. The node then writes its own IP address as the Parcel Reply source and writes the source of the Parcel Probe as the Parcel Reply destination (for IPv4 Parcel Probes, the node writes the Parcel Reply addresses as IPv4-Compatible IPv6 addresses ). The node next copies as much of the leading portion of the Parcel Probe (beginning with the IP header) as possible into the "packet in error" field without causing the Parcel Reply to exceed 512 octets in length, then calculates the ICMPv6 header checksum. Since IPv6 packets cannot traverse IPv4 paths, and since middleboxes often filter ICMPv6 messages as they traverse IPv6 paths, the node next wraps the Parcel Reply in UDP/IP headers of the correct IP version with the IP source and destination addresses copied from the Parcel Reply and with UDP port numbers set to the UDP port number for OMNI . In the process, the node either calculates or omits the UDP checksum as appropriate and (for IPv4) clears the DF bit. The node finally sends the prepared Parcel Reply to the original source of the probe. After sending a Parcel Probe the original source may therefore receive a UDP/IP encapsulated Parcel Reply (see above) and/or an upper layer protocol probe reply. If the source receives a Parcel Reply, it first verifies the checksum(s) then matches the enclosed PTB message with the original Parcel Probe by examining the Nonce-2 field echoed in the ICMPv6 "packet in error" field containing the leading portion of the probe. If PTB does not match, the source discards the Parcel Reply; otherwise, it continues to process. If the Parcel Reply MTU is 0, the source marks the path as "parcels not supported; otherwise, it marks the path as "parcels supported" and also records the MTU value as the parcel path MTU for the forward path to this destination. (Note that this size may be larger than the maximum-sized singleton jumbogram that can traverse the path.) After receiving a positive Parcel Reply, the original source can continue sending IP parcels addressed to the final destination up to the size of the parcel path MTU; any upper layer protocol probe replies will determine the maximum segment size that can be included in the parcel as an upper layer consideration. After receiving a negative Parcel Reply (or no reply) the original source should refrain from sending parcels. In both cases, the original source should then periodically re-initiate Parcel Path Qualification as long as it continues to prefer to use the IP parcel service. If at any time performance appears to degrade, the original source should cease sending IP parcels and/or reduce the size of the parcels it sends. Nodes can also use this Parcel Path Qualification procedure to qualify the path for ordinary IP jumbograms simply by setting Nonce-1 to the value '11111110' and formatting the probe body as an ordinary jumbogram no larger than the maximum size that can be represented in the 32-bit Jumbo Payload Length. Nodes that forward the (Jumbogram) Parcel Probe will recognize the Nonce-1 value as an indication that the probe is a true Jumbogram (i.e., and not a parcel). The node then sets PMTU to the largest possible Jumbogram size and forwards the probe to the next hop. Nodes that return (Jumbogram) Parcel Replys will then return the resulting PMTU value. This especially implies the largest possible Jumbogram size may be less than the largest possible parcel size, since forwarding nodes can sub-divide parcels but cannot sub-divide singleton Jumbograms. Note: when a Parcel Probe forwarded into an ingress OMNI interface is broken into sub-parcels, each sub-parcel includes its own copy of the Parcel Probe header. When multiple sub-parcels of the same Parcel Probe arrive at an egress OMNI interface, the interface optionally re-combines the sub-parcels while retaining the Parcel Probe header. It is therefore possible that a single Parcel Probe with multiple upper layer protocol segments could generate multiple Parcel Replys. Note: The original source includes Nonce-1 and Check fields as the first two octets of Parcel Probes in case a router on the path overwrites the values in a wayward attempt to implement . Parcel Probe recipients should therefore regard a Nonce-1 value other than '11111111' or '11111110' as an indication that the field was either intentionally or accidentally altered by a previous hop node that does not recognize parcels. Note: The MTU value returned in a Parcel Reply determines only the maximum IP parcel size for the path, while the maximum upper layer protocol segment size may be significantly smaller. The upper layer protocol segment size is instead determined separately according to any upper layer protocol probing. Note: When the OMNI interface of an ingress middlebox receives a Parcel Probe with PMTU larger than 64KB, it can optionally leave PMTU unchanged (i.e., if it intends to support parcel subdivision internally) or rewrite PMTU to 64KB to disable adaptation layer parcel sub-division. Regardless of the decision taken by the ingress middlebox, correct behavior will be observed by the final destination whether or not the egress middlebox elects to recombine sub-parcels. Note: If a router or final destination receives a Parcel Probe but does not recognize the parcel construct, it simply drops the probe. The original source will then deem the probe as lost and parcels cannot be used.

The {TCP,UDP}/IP header plus each segment of a (multi-segment) IP parcel includes its own integrity check. This means that IP parcels can support stronger and more discrete integrity checks for the same amount of upper layer protocol data compared to an ordinary IP packet or Jumbogram. The {TCP/UDP} header integrity checks can be verified at each hop to ensure that parcels with errored headers are dropped to avoid mis-delivery (nodes that set and verify TCP parcel header checksums must honor the sequence number discipline discussed in ). The header and per-segment integrity checks must then be verified at the final destination. IP parcels can range in length from as small as only the {TCP,UDP}/IP headers plus a single checksum with a segment of length 1 to as large as the headers plus (64 * (65535 minus headers)) octets. Although 32-bit link layer integrity checks provide sufficient protection for contiguous data blocks up to approximately 9KB, reliance on link-layer integrity checks may be inadvisable for links with significantly larger MTUs and may not be possible at all for links such as tunnels over IPv4 that invoke fragmentation. Moreover, the segment contents of a received parcel may arrive in an incomplete and/or rearranged order with respect to their original packaging. Lower layer protocol entities always calculate and verify the {TCP,UDP}/IP parcel header checksum at their layer, since an errored header could result in mis-delivery to the wrong upper layer protocol entity. If the lower layer protocol entity of the destination detects an incorrect {TCP,UDP}/IP checksum it discards the entire IP parcel. To support the parcel header checksum calculation, lower layer protocol entities use modified versions of the {TCP,UDP}/IPv4 "pseudo-header" found in , or the {TCP,UDP}/IPv6 "pseudo-header" found in Section 8.1 of . Note that while the contents of the two IP protocol version-specific pseudo-headers beyond the address fields are the same, the order in which the contents are arranged differs and must be honored according to the specific IP protocol version as shown in . This allows for maximum reuse of widely deployed code while ensuring interoperability.

where: Source Address is the 4-octet IPv4 or 16-octet IPv6 source address of the prepared parcel. Destination Address is the 4-octet IPv4 or 16-octet IPv6 destination address of the prepared parcel. zero encodes the constant value '0'. Next Header is the IP protocol number corresponding to the upper layer protocol, i.e., TCP or UDP. Segment Length is the value that appears in the IPv4 Total Length or IPv6 Payload Length field of the prepared parcel. Nsegs is the 1-octet number of segments included, and must contain a number between 1 and 64 (this is the same value that appears in the Jumbo Payload Option). Upper-Layer Packet Length is the 3-octet length of the {TCP,UDP} header plus TCP data (this value can be derived from the Jumbo Payload Length by subtracting the IPv4 header length for IPv4 or IPv6 extension header length for IPv6). Upper layer protocol entities use socket options to coordinate per-segment checksum processing with lower layers. If the upper layer sets a SO_NO_CHECK(TX) socket option, the upper layer is responsible for supplying per-segment checksums on transmission and the lower layer forwards the IP parcel to the next hop without further processing; otherwise, the lower layer calculates the per-segment checksums before forwarding. If the upper layer sets a SO_NO_CHECK(RX) socket option, the upper layer is responsible for verifying per-segment checksums on reception and the lower layer delivers each received parcel body to the upper layer without further processing; otherwise, the lower layer verifies the per-segment parcel checksums before delivering. When the upper layer protocol entity of the source sends a parcel body to lower layers, it prepends a block of Nsegs 2-octet Checksum fields and includes a 4-octet Sequence Number field with each TCP non-first segment. If the SO_NO_CHECK(TX) socket option is set, the upper layer protocol either calculates each segment checksum and writes the value into the checksum field (with '0' values written as 'ffff') or writes the value '0' to disable checksums for specific segments. If the SO_NO_CHECK(TX) socket options is clear, the upper layer instead writes the value '0' to disable or any non-zero value to enable checksums for specific segments. When the lower layer protocol entity of the source receives the parcel body from upper layers, if the SO_NO_CHECK(TX) socket option is set the lower layer appends the {TCP,UDP}/IP headers and forwards the parcel to the next hop without further processing. If the SO_NO_CHECK(TX) socket option is clear, the lower layer instead calculates the checksum for each segment with a non-zero value in the Checksum field and overwrites the calculated value into the Checksum field (with '0' values written as 'ffff'). When the lower layer protocol entity of the destination receives a parcel from the source, if the SO_NO_CHECK(RX) socket option is set the lower layer delivers the parcel body to the upper layer without further processing, and the upper layer is responsible for per-segment checksum verification. If the SO_NO_CHECK(RX) socket option is clear, the lower layer instead calculates the checksum for each segment with a non-zero value in the Checksum field and overwrites the field with the value '1' if the checksum is correct or any other non-zero value if the checksum is incorrect. The lower layer then delivers the parcel body (beginning with the Checksum block) to the upper layer, which will see the value '0' for checksums disabled, '1' for checksum correct or any other value for checksum incorrect in each segment Checksum. Note: The checksum block itself is intentionally omitted from the IP Parcel {TCP,UDP} header checksum calculation. This permits destinations to accept as many intact segments as possible from received parcels with checksum block bit errors, whereas the entire parcel would need to be discarded if the header checksum also covered the checksum block. Note: Implementations may provide a configuration option that allows lower layers to deliver the actual checksum received in an errored parcel segment to upper layers instead of a random value other than '0' or '1', e.g., for logging purposes. If so, the lower layer should rewrite actual '1' checksums to 'ffff' to allow upper layers to discern correct from errored checksums.

Section 3 of provides a list of certain conditions to be considered as errors. In particular: error: IPv6 Payload Length != 0 and Jumbo Payload option present error: Jumbo Payload option present and Jumbo Payload Length < 65,536 Implementations that obey this specification ignore these conditions and do not consider them as errors.

By defining a new IPv4 Jumbo Payload option, this document also implicitly enables a true IPv4 jumbogram service defined as an IPv4 packet with a Jumbo Payload option included and with Total Length set to 0. All other aspects of IPv4 jumbograms are the same as for IPv6 jumbograms .

Common widely-deployed implementations include services such as TCP Segmentation Offload (TSO) and Generic Segmentation/Receive Offload (GSO/GRO). These services support a robust (but not standardized) service that has been shown to improve performance in many instances. UDP/IPv4 parcels have been implemented in the linux-5.10.67 kernel and ION-DTN ion-open-source-4.1.0 source distributions. Patch distribution found at: "https://github.com/fltemplin/ip-parcels.git". Testing with a single-threaded receiver has shown that including increasing numbers of segments in a single parcel produces modest performance gains over fewer numbers of segments due to more efficient packaging and reduced system calls/interrupts. Since performance is strongly bounded by single segment receiver processing time (with larger segments producing dramatic performance increases), it is expected that parcels with increasing numbers of segments will provide a performance multiplier on multi-threaded receivers running in parallel processing environments.

The IANA is instructed to change the "MTUP - MTU Probe" entry in the 'ip option numbers' registry to the "JUMBO - IPv4 Jumbo Payload" option. The Copy and Class fields must both be set to 0, and the Number and Value fields must both be set to '11'. The reference must be changed to this document [RFCXXXX].

Original sources match the Nonce values in received Parcel Replys with their corresponding Parcel Probes. If the values match, the reply is likely an authentic response to the probe. In environments where stronger authentication is necessary, nodes that send Parcel Replys can apply the message authentication services of OMNI and Automatic Extended Route Optimization (AERO) . Multi-layer security solutions may be necessary to ensure confidentiality, integrity and availability in some environments.

This work was inspired by ongoing AERO/OMNI/DTN investigations. The concepts were further motivated through discussions on the intarea and 6man lists. A considerable body of work over recent years has produced useful "segmentation offload" facilities available in widely-deployed implementations.

Historic and current-day data links configure Maximum Transmission Units (MTUs) that are far smaller than the desired state for the future of IP parcel transmission. When the first Ethernet data links were deployed many decades ago, their 1500 octet MTU set a strong precedent that was widely adopted. This same size now appears as the predominant MTU limit for most paths in the Internet today, although modern link deployments with larger MTUs up to 9KB have begun to emerge. In the late 1980's, the Fiber Distributed Data Interface (FDDI) standard defined a new link type with MTU slightly larger than 4500 octets. The goal of the larger MTU was to increase performance by a factor of 10 over the ubiquitous 10Mbps and 1500 octet MTU Ethernet technologies of the time. Many factors including a failure to harmonize MTU diversity and an Ethernet performance increase to 100Mbps led to the demise of FDDI. Moving into the next decade, the 1990's saw new initiatives including ATM/AAL5 (9KB MTU) and HiPPI (64KB MTU) which offered high-speed data link alternatives with larger MTUs but again the inability to harmonize diversity derailed their momentum. By the end of the 1990s and leading into 2000's, emergence of the 1Gbps and faster Ethernet performance levels seen today has obscured the fact that the modern Internet of the 21st century is still operating with 20th century MTUs! To bridge this gap, increased OMNI interface deployment in the near future will provide an unlimited MTU virtual link type that can pass IP parcels over paths that traverse traditional data links with small MTUs. Experiments have shown that (single-threaded) receive-side performance is bounded by upper layer protocol segment size, with performance increasing in direct proportion with segment size. Experiments have also shown that (single-threaded) performance increases moderately by including larger numbers of segments per parcel. However, parallel receive-side processing will provide performance multiplier benefits since the multiple segments that arrive in a single parcel can be processed simultaneously instead of serially. But, the true power of IP parcels will become evident as future parcel-capable links with extremely large MTUs begin to emerge. These links will provide MTUs far in excess of 64KB (up to as large as 4MB). With such large MTUs, the traditional CRC-32 (or even CRC-64) error checking with errored packet discard discipline will no longer apply for large parcels. Instead, parcels larger than a link-specific threshold will include Forward Error Correction (FEC) codes so that errored parcels can be repaired at the receiver's data link layer then delivered to upper layers rather than being discarded and triggering retransmission of large amounts of data. Even if the FEC repairs are incomplete or imperfect, all parcels can still be delivered to upper layers where the individual segment checksums will detect and discard any damaged data not repaired by lower layers. These new "super-links" will most likely occur in the network edges (e.g., high-speed data centers) and would not be expected to occur in the middle of the Internet. (However, some space-domain links that extend over enormous distances may also benefit.) For this reason, a common use case will include parcel-capable super-links in the edge networks of both parties of an end-to-end session with an OMNI link connecting the two over wide area Internetworks. Small- to medium-sized IP parcels over OMNI links will already provide considerable performance benefits for wide-area end-to-end communications while truly large IP parcels over super-links can provide boundless increases for localized bulk transfers in edge networks or for deep space long haul transmissions. The ability to grow and adapt without practical bound enabled by IP parcels will inevitably encourage new data link development leading to future innovations in new markets that will revolutionize the Internet. Until these new links begin to emerge, however, parcels will already provide a tremendous benefit to end systems by allowing applications to send and receive segment buffers larger than 65535 octets in a single system call. By expanding the current operating system call data copy limit from its current 16-bit length to a 32-bit length, applications will be able to send and receive maximum-length parcel buffers even if lower layers need to break them into multiple parcels to fit within the underlying interface MTU. For applications such as the Delay Tolerant Networking (DTN) Bundle Protocol , this will allow applications to send and receive entire large DTN bundles in a single system call.