| Internet-Draft | SHIP for Edge Networks | April 2023 |
| Song | Expires 14 October 2023 | [Page] |
To mitigate the IPv6 header overhead and improve the scalability and performance in edge networks, this draft proposes to use short hierarchical IP addresses excluding the network prefix within edge networks. An edge network can be further organized into a hierarchical architecture containing one or more levels of networks. While each end node only needs to keep a short address suffix as its identifier, the border routers for each hierarchical level are responsible for address augmenting and pruning when a packet leaves or enter a lower level network. Specifically, the top-level border routers of an edge network convert the internal IP header to and from the standard IPv6 header. This draft presents an incrementally deployable scheme allowing packet header to be effectively compressed in edge networks without affecting the network interoperability. Simplifying both network data plane and control plane, the SHIP architecture is suitable for any types of edge networks, especially when low latency, high performance, and high bandwidth efficiency are required.¶
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 [RFC2119][RFC8174] when, and only when, they appear in all capitals, as shown here.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 14 October 2023.¶
Copyright (c) 2023 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Internet of Things (IoT) and 5G introduce to the Internet a huge number of addressable entities (e.g., sensors, machines, vehicles, and robots). The transition to IPv6 is inevitable. While the 128-bit address of IPv6 was considered large enough and future-proof, the long IP addresses inflate the packet header size. 80% of a basic IPv6 header is consumed by addresses.¶
In IoT networks, thing-to-thing communication through wireless connections is dominant, which presents several distinct characteristics. (1) The communication pattern is often frequent short-message exchanges (e.g., industry robots and networked vehicles). (2) The communication is usually energy sensitive (e.g., battery-powered sensors). (3) The communication often requires low latency (e.g., industry control). (4) The precious wireless channels demand high bandwidth utilization (e.g., ZigBee, Bluetooth, Wi-Fi, and 5G). These characteristics render a large header overhead unfavorable and even prohibitive.¶
The address overhead also takes its toll on Data Center Networks (DCN), especially when large scale containers are deployed, the east-west traffic is dominant, and the prevailing communications are comprised of short messages (e.g., key-value pairs) and conducted through virtual switches.¶
In IoT and DCN, since most communications happen between adjacent and related entities, it is a good practice to locally confine communication, computing, and storage due to performance, efficiency, and security considerations, as advocated by Edge Computing. Such a communication pattern provides an opportunity to mitigate the IPv6 header overhead problem due to the long addresses.¶
When an IPv6 address block is allocated to an edge network, all the entities in the edge network share the same address prefix. When these entities communicate with each other, they can ignore the common prefix. In fact, they don not even need to know the common prefix. Only when they need to communicate with entities outside of the edge network, the full addresses are needed. Even in this case, the entities in the edge network still do not need to know the prefix. It is sufficient for the gateway routers at the network border to manipulate the addresses (i.e., augmenting or pruning the address) to meet the addressing requirement.¶
Following this line of thought, an edge network can be further partitioned into multiple hierarchical levels, which support flexible sub-networking. The result is that an end entity needs to maintain an even shorter address as its identifier. For communication crossing network levels, the address manipulation is done at each gateway router on the path recursively.¶
In this draft, we define an edge network as a stub network which does not support traffic transit service. The stub network is assigned an IPv6 address block (i.e., a prefix). In this sense, a data center network in cloud can also be considered as an edge network. An edge network usually falls under a single network administration domain.¶
The address block assigned to an edge network is identified by a prefix P with the length of L < 128 bits. The remaining S = 128-L bits can be used to assign addresses to the entities in this network. A key observation is: the entities in this network do not need to be aware of P's length and value at all. We can further partition the edge network into multiple hierarchical levels, making a tree architecture. The root represents the entire edge network. Each other node represents a lower level network occupying a sub address space owned by its parent node. A leaf node represents a lowest level network. We name the root level network the L_0 network. Its children are all L_1 networks, and so on so forth. In other words, the network level is the depth of the corresponding node in the tree.¶
The network hierarchy partitions the S-bit address into multiple sections. Assume an entity is in an L_n network. The S-bit address is partitioned into n+1 sections. The entity only needs to keep the last section of the S-bit address as its ID. The gateway routers for each level of network maintain one section of the S-bit address. Specifically, the gateway routers of L_i (i>0) keep the i-th section of the S-bit address, and the gateway routers of L_0 keep the assigned IPv6 address block prefix P.¶
Figure 1 shows an edge network example, in which are three network levels. The edge network A is assigned a 96-bit IPv6 address prefix (2001:0db8:ac10:fe01::0001), which means it owns a 32-bit address space. In this space, two L_1 networks are created: B with a 16-bit prefix (0xaaaa) and C with a 24-bit prefix (0xcccccc). Note that the prefixes at the same level must not overlap in order to guarantee entities in the edge network are uniquely addressable. Network B contains two entities x and y, and Network C contains one entity z. In network B, an L_2 network C is further created with a 8-bit prefix (0xbb). In this example, an entity in C or D (e.g., m and z) only need to own a 8-bit address, an entity in B but not in D (e.g., x and y) needs to own a 16-bit address, and an entity in A but not in B and C needs to own a 32-bit address. In this way, each entity in A still logically owns a unique IPv6 address (e.g., the IPv6 address of the entity m in D with ID of 5 is 2001:0db8:ac10:fe01::0001:aaaa:bb05), although the entity m is only aware of its local ID (0x05).¶
+------+
| L0:A | 2001:0db8:ac10:fe01::0001/96
+-+--+-+
| |
+------+ +-------+
| |
+---+--+ +--+---+
(x,y) | L1:B | aaaa/16 | L1:C | cccccc/24
+---+--+ +------+
| (z)
+---+--+
(m) | L2:D | bb/8
+------+
The edge networks adopting the short and variable size address scheme need a new type of IP header, which is referred as IPvn in this draft. Apart from the IP version, the major difference between IPvn and IPv6 headers is the address fields. IPvn replaces IPv6's 128-bit source address field and 128-bit destination address field with the four fields shown in Figure 2.¶
+------------+------------+
| SAL | DAL |
+------------+------------+
| SA (variable length) |
+-------------------------+
| DA (variable length) |
+-------------------------+
The Source Address Length (SAL) and the Destination Address Length (DAL) fields have fixed length, while the Source Address (SA) and the Destination Address (DA) fields are of variable length. To simplify the implementation, SA and DA are preferred to be byte-aligned. It is possible to define the length of address in the unit of byte, nibble, or bit. Each has its own pros and cons. The unit of byte can help reduce the size of the SAL/DAL but results in coarse network granularity which might be inefficient in address allocation. For example, a 3-bit SAL/DAL is enough to encode 8 possible address lengths (one to eight bytes) for networks. In this design, each higher level network's address space expands 256 times. On the other extreme, the unit of bit allows fine network granularity but requires more space for SAL/DAL. For example, 6-bit SAL and DAL can support an address length up to 64 bits (8 bytes) and each higher level network is only twice larger.¶
With a few bits, it is also possible to design a more sophisticated encoding scheme that supports variable address length steps and adapts to the ideal network sizes at different levels.¶
Assuming SA and DA are 2 bytes each, and SAL and DAL are 4 bits each, the address fields are only 5 bytes in total. Comparing to IPv6, the size of the address fields is reduced by 84%.¶
In the edge network hierarchy, each network has one or more Level Gateway Routers (LGR) which are responsible for forwarding packets in or out of this network. The LGRs are the only interface between a network and its parent network.¶
A network can be in a single L2 domain, which means all the entities in this network (excluding those in its child networks) and all the network devices (including the LGRs to the parent network and the child networks) are L2 reachable. A network can also be a pure L3 network in which no L2 device is allowed. Each entity in a network is directly connected to either an LGR or some internal routers named Intra-Level Router (ILR) which is solely responsible for packet forwarding within the network. In this case, the entities need to partially participate in the routing process (e.g., advertising its address).¶
The scale of an intra-level network can be used to guide the L2/L3 selection. Small networks prefer the L2-based solution and large networks prefer the L3-based solution. In the higher level networks (e.g., closer to the top level network or the tree root), since the number of entities is usually small, it is free to choose between L2 or L3-based solution. The leaf level networks are usually L2-based for simplicity.¶
Unlike in IPv4 and IPv6 networks, the address related fields in IPvn header can be modified by LGRs. An LGR of a network keeps a prefix that can augment the SAs from this network to an address outside of this network. If an LGR needs to forward an internal packet outside (i.e., DAL > SAL), it augments the packet's SA and updates its SAL accordingly. Reversely, if an LGR receives a packet from the parent network destined for the child network for which it serves as a gateway (i.e., the parent network prefix matches the DA's prefix), it strips off the parent network prefix from the packet's DA and updates its DAL accordingly.¶
In contrast, within an L3-based level network, ILRs do not modify the address fields. An ILR can decide the packet forwarding direction by examining the DAL. If DAL > SAL, the packet needs to be forwarded to an LGR of this network; otherwise, the packet needs to be forwarded within the current network, and possibly into a lower-level child network.¶
An LGR of the top-level network (i.e., the L0 network) is special. In addition to the address manipulation, it is also responsible for converting the IPvn header to and from the standard IPv6 header to support the Internet interoperability. We name such a router IP Translator (IPT).¶
We use the edge network shown in Figure 1 to illustrate some packet forwarding examples. The details for the involved entities are summarized in Figure 3. In the IPvn packet header, we use 4 bits to encode the address length. In particular, 0b0000 is used to indicate the address is 16 bytes long (i.e., a complete IPv6 address).¶
+------+------+------+-----+-------+------------+ |Entity| ID |Length|Level|Network| Prefix | +------+------+------+-----+-------+------------+ | x |0x0001| | | | | +------+------|2bytes| 1 | B |0xaaaa/16 | | y |0x0002| | | | | +------+------+------+-----+-------+------------+ | z |0x01 |1byte | 1 | C |0xcccccc/24 | +------+------+------+-----+-------+------------+ | m |0x08 |1byte | 2 | D |0xbb/8 | +------+------+------+-----+-------+------------+
The first example in Figure 4 shows how packets are forwarded from x to y within the same network B. In this case, the source address and destination address have the same length. The packets only pass through an ILR which does not change the address fields.¶
+-----------------+ +-----------------+ +-----------------+
| IPvn Header | | IPvn Header | | IPvn Header |
+--------+--------+ +--------+--------+ +--------+--------+
|SAL:0x2 |DAL:0x2 | |SAL:0x2 |DAL:0x2 | |SAL:0x2 |DAL:0x2 |
+--------+--------+ +--------+--------+ +--------+--------+
|SA: 0x0001 | |SA: 0x0001 | |SA: 0x0001 |
+-----------------+ +-----------------+ +-----------------+
|DA: 0x0002 | |DA: 0x0002 | |DA: 0x0002 |
+-----------------+ +-----------------+ +-----------------+
Entity x ------> ILR in B ------> Entity y
The second example in Figure 5 shows how packets are forwarded from x in B to z in C. At LGR of B, the source address is augmented, and at the LGR of C, the destination address is pruned. Since x and z's nearest common ancestor network is A, so the packets never need to leave network A, so A's prefix is oblivious throughout the communication.¶
+---------------+ +---------------+ +---------------+ +---------------+
| IPvn Header | | IPvn Header | | IPvn Header | | IPvn Header |
+-------+-------+ +-------+-------+ +-------+-------+ +-------+-------+
|SAL:0x2|DAL:0x4| |SAL:0x4|DAL:0x4| |SAL:0x4|DAL:0x1| |SAL:0x4|DAL:0x1|
+-------+-------+ +-------+-------+ +-------+-------+ +-------+-------+
|SA: 0x0001 | |SA: 0xaaaa0001 | |SA: 0xaaaa0001 | |SA: 0xaaaa0001 |
+---------------+ +---------------+ +---------------+ +---------------+
|DA: 0xcccccc01 | |DA: 0xcccccc01 | |DA: 0x01 | |DA: 0x01 |
+---------------+ +---------------+ +---------------+ +---------------+
Entity x -----> LGR of B -----> LGR of C -----> Entity z
The last example in Figure 6 shows how packets are forwarded from x in B to a host in IPv6 domain. In the IPT of A, the IPvn header is converted to an IPv6 header.¶
+---------------+ +---------------+ +---------------+ +---------------+
| IPvn Header | | IPvn Header | | IPv6 Header | | IPv6 Header |
+-------+-------+ +-------+-------+ +---------------+ +---------------+
|SAL:0x2|DAL:0x0| |SAL:0x4|DAL:0x0| |SA: 2001:0db8 | |SA: 2001:0db8 |
+-------+-------+ +-------+-------+ | ac10:fe01: | | ac10:fe01: |
|SA: 0x0001 | |SA: 0xaaaa0001 | | 0000:0001: | | 0000:0001: |
+---------------+ +---------------+ | aaaa:0001 | | aaaa:0001 |
|DA: 2001:0db8: | |DA: 2001:0db8: | +---------------+ +---------------+
| 85a3:0000: | | 85a3:0000: | |DA: 2001:0db8: | |DA: 2001:0db8: |
| 0000:8a2e: | | 0000:8a2e: | | 85a3:0000: | | 85a3:0000: |
| 0370:7334 | | 0370:7334 | | 0000:8a2e: | | 0000:8a2e: |
+---------------+ +---------------+ | 0370:7334 | | 0370:7334 |
+---------------+ +---------------+
Entity x -----> LGR of B -----> IPT of A -----> Entity n
Within the edge networks where IPvn is applied, all the control plane functions and protocols need to be modified or redesigned due to the hierarchical network architecture of IPvn. Fortunately, the updates are often incremental and the results are usually simpler than their counterparts in IPv4 and IPv6. We briefly discuss a few essential protocols that enable the operation of IPvn.¶
To expand the address space of the edge network, the IPT of the edge network can also support functions similar to NAT. In this case, the edge network is assigned one or more public IPv4/IPv6 addresses. The entities in IPvn domain use private addresses. The IPT maintains the mapping table between the private address and public address.¶
Although the motivation of this draft is to support shorter address (i.e., smaller L3 header overhead) in edge networks, it is worth noting that the scheme allows the addresses to be extend to arbitrary length, even longer than 128bits. In that case, the address space of the IPvn network can be greater than that of IPv6 and the entire IPv6 network can be considered an edge network of the IPvn network. This scenario should be considered when specifying the address fields of IPvn.¶
IPv6 header compression schemes have been specified for some particular low power IoT networks such as 6loWPAN ([RFC6282]) and LPWAN ([RFC8724]). These networks feature low data rate and are insensitive to latency, however, due to the low power constraint, they are extremely sensitive to bandwidth efficiency. Therefore, they adopt the context-based compression schemes which, while needing extra storage and computation, can reduce the header overhead to the utmost extend.¶
In contrast, SHIP is context-less and independent to the edge network type. Hence, SHIP is free from the packet-based compression/decompression process and the context maintenance, making it suitable for high bandwidth and low latency communications. Also, SHIP provides a hierarchical network architecture which allows better network manageability and isolation.¶
The current proposal only concerns the address part of the IPv6 packet header. In edge networks and for particular applications, the context-less field eliding and reduction on the other non-essential IPv6 header fields are possible to further reduce the header overhead while maintaining the high performance.¶
Below is a list of potential use cases in addition to the DCN discussed in Section 1 which can appreciate the unique property of SHIP.¶
The SHIP addressing scheme and architecture allow a securer edge network. The IPTs and LGRs naturally support the access control.¶
The proposal requires to use a new IP version and define a new IP header which can be converted to/from an equivalent IPv6 header.¶
We acknowledge the technical contributions, suggestions and comments from Yingzhe Qu, Zhaobo Zhang, James Guichard, Toerless Eckert, Stewart Bryant, Michael McBride, Adnan Rashid, Alexander Pelov, Michael Richardson, Pascal Thubert, Uma Chunduri, Kerry Lynn, and many others.¶