This is the first post in a series that will discuss a new layer in Internet architecture known as the adaptation layer, including the related cost-benefit implications of its introduction.
The concept of the Internet layered architecture had its foundations in the OSI 7-layer model but has more recently been described as an ‘hourglass model’ with all upper- and lower-layer functions coordinating via the Internet Protocol (IP) layer as a narrow conduit.
Upper layer protocols such as TCP and UDP use IP layer packetization services for forwarding packets over underlying data-link interfaces, which may include a heterogeneous mix of technologies such as Ethernet, Wi-Fi, and 4G/5G cellular. As per the original OSI terminology, the IP layer is still known in modern architecture as Layer 3 (L3) and the data link layer as Layer 2 (L2).
Pushing the boundaries of packet forwarding
In the current Internet architecture, users are encouraged to restrict packet sizes to no larger than the smallest link in the path from the source to the destination — a characteristic widely known as the Path Maximum Transmission Unit (Path MTU). But, this places unnecessary limits on the architecture that impedes growth.
IP fragmentation can assist with breaking packets into smaller pieces (fragments) so that the resulting pieces can pass through a link with a smaller MTU than the original packet size. Although using the IP layer fragmentation and reassembly service is discouraged and even considered ‘harmful’ by some, this process can be quite effective. Some studies have shown applications often see greater performance by sending packets larger (sometimes much larger) than the Path MTU, while intentionally invoking IP fragmentation and reassembly as necessary. This has not gone unnoticed by network operators who see the need for a new layer in the Internet architecture known as the ‘adaptation layer’.
The adaptation layer appears below the IP layer (L3) but above the data link layer (L2). No layer number is associated with the adaptation layer since numbering is an artifact of the historic OSI reference model and need not be carried forward into future architecture. It is therefore sufficient to simply consider the adaptation layer as ‘the layer between L3 and L2’ manifested by the Overlay Multilink Network Interface (OMNI).
OMNI adaption layer for a future Internet
The OMNI includes an OMNI Adaptation Layer (OAL) based on IPv6 encapsulation, fragmentation, and reassembly. By invoking fragmentation and reassembly at the OAL layer, the IP layer can forward packets larger than the Path MTU and any fragmentation and reassembly will be invisible to IP since it occurs at a lower layer in the architecture.
The OAL concept is not new. In fact, it was the subject of an industry-wide investigation in the 1990s as part of the Asynchronous Transfer Mode (ATM) grand-scale experiment.
ATM included ATM Adaptation Layer 5 (AAL5) as a layer between IP and the underlying data links which were only capable of transporting 53 octet cells. However, the ATM experiment showed that AAL5 could forward large IP packets (up to 9,180 octets) over data links with such small cell sizes using an adaptation layer segmentation and reassembly service.
The proposal to include an OAL for the future Internet differs from AAL5 in two ways:
- It intentionally applies to heterogeneous data links that could have diverse MTUs (as opposed to tiny cell sizes).
- Unlike L2 ATM gear, the service model for heterogeneous data links is best-effort only, with the loss of individual fragments possible.
The OAL encapsulates original IP packets in IPv6 encapsulation headers, inserts a fragment header, then applies IPv6 fragmentation if necessary. The OAL can apply fragmentation for packets as large as (216-1) octets using a variable-length ‘cell size’ determined by the path MTU. The OAL then forwards the fragments over underlying data-link interfaces as the first hop in a path that may span one or more Internetworks.
Since the Internet service model is ‘best-effort’, it is possible that one or more fragments of an individual packet could be lost, making the loss unit smaller than the retransmission unit. This can lead to cascading failures when the original source retransmits the entire original packet (which could be quite large) only to have the entire retransmission lost again if even just one or a few small fragments are lost. The OAL, therefore, provides two complementary mitigations.
First, the OAL considers the adaptation layer source and destination as the endpoints of a virtual link. The OAL source and destination may be co-located with the original IP packet source and destination application endpoints, but often they are located somewhere in the middle of the network path.
As intermediate endpoints, the virtual link between the OAL source and destination can apply a (virtual) link-layer Automatic Repeat Request (ARQ) service so that individual lost fragments can be quickly retransmitted before the application endpoints negotiate retransmission of the entire original IP packet. This ‘just-in-time’ retransmission can often suppress cascading failures due to application endpoint retransmissions. It is based on a new Internet Control Message Protocol (ICMP) message type known as the ‘Fragmentation Report’ that allows the OAL destination to inform the OAL source of individual missing fragments.
The OAL source then satisfies the partial reassembly by retransmitting the missing fragments.
Second, since the fragment retransmission service is best-effort and with a small convergence window, an additional mitigation step is necessary to handle cases when excessive fragment loss occurs. For example, if the OAL destination sees more than just an occasional lost fragment it may require some means to inform the original source to reduce the size of the packets it is sending.
For this purpose, a new type of ICMP Packet Too Big (PTB) message known as a ‘PTB soft error’ is defined. The OAL destination can issue a PTB soft error to request the original source to begin sending smaller packets with fewer fragments to improve performance within the current fragment loss profile. The original source can then later resume sending large packets if the PTB soft errors subside.
Tuning packet sizes to match changing network conditions
Together, the PTB soft errors and Fragmentation Reports provide a useful best-effort reliability service to maximize throughput while minimizing loss. The service goal is to allow an original source to send the largest possible packets within current fragment loss and retransmission constraints and thereby dynamically tune its packet sizes to match changing network conditions.
In future posts, I will discuss other important aspects of the OAL including integrity, efficiency, and security.
Fred Templin is an Internet networking research engineer working in the industry since 1986, where he has been deeply involved in the evolution of the Internet Protocol over Ethernet, FDDI, ATM, wireless and other data link and network layer technologies.
The views expressed by the authors of this blog are their own and do not necessarily reflect the views of APNIC. Please note a Code of Conduct applies to this blog.