Huawei Technologies Co., Ltd

Q14, Huawei Campus, No.156 Beiqing Road Hai-Dian District, Beijing, 100095 China jiangsheng@huawei.com

Department of Computer Science University of Auckland PB 92019 Auckland 1142 New Zealand brian.e.carpenter@gmail.com

Cisco Systems

Building D, 45 Allee des Ormes Mougins 06250 France mbehring@cisco.com

IRTF Network Management This document provides a problem statement and general gap analysis for an IP-based Autonomic Network that is mainly based on distributed network devices. The document provides background by reviewing the current status of autonomic aspects of IP networks and the extent to which current network management depends on centralization and human administrators. Finally, the document outlines the general features that are missing from current network abilities and are needed in the ideal Autonomic Network concept. This document is a product of the IRTF's Network Management Research Group.

The general goals and relevant definitions for Autonomic Networking are discussed in . In summary, the fundamental goal of an Autonomic Network is self-management, including self-configuration, self-optimization, self-healing, and self-protection. Whereas interior gateway routing protocols such as OSPF and IS-IS largely exhibit these properties, most other aspects of networking require top-down configuration, often involving human administrators and a considerable degree of centralization. In essence, Autonomic Networking is putting all network configurations onto the same footing as routing, limiting manual or database-driven configuration to an essential minimum. It should be noted that this is highly unlikely to eliminate the need for human administrators, because many of their essential tasks will remain. The idea is to eliminate tedious and error-prone tasks, for example, manual calculations, cross-checking between two different configuration files, or tedious data entry. Higher-level operational tasks, and complex troubleshooting, will remain to be done by humans. This document represents the consensus of the IRTF's Network Management Research Group (NMRG). It first provides background by identifying examples of partial autonomic behavior in the Internet and by describing important areas of non-autonomic behavior. Based on these observations, it then describes missing general mechanisms that would allow autonomic behaviors to be added throughout the Internet.

The terminology defined in is used in this document.

This section discusses the history and current status of automatic or autonomic operations in various aspects of network configuration, in order to establish a baseline for the gap analysis. In particular, routing protocols already contain elements of autonomic processes, such as information exchange and state synchronization.

For many years, there was no alternative to completely manual and static management of IP addresses and their prefixes. Once a site had received an IPv4 address assignment (usually a Class C /24 or Class B /16, and rarely a Class A /8), it was a matter of paper-and-pencil design of the subnet plan (if relevant) and the addressing plan itself. Subnet prefixes were manually configured into routers, and /32 addresses were assigned administratively to individual host computers and configured manually by system administrators. Records were typically kept in a plain text file or a simple spreadsheet. Clearly, this method was clumsy and error-prone as soon as a site had more than a few tens of hosts, but it had to be used until DHCP became a viable solution during the second half of the 1990s. DHCP made it possible to avoid manual configuration of individual hosts (except, in many deployments, for a small number of servers configured with static addresses). Even so, prefixes had to be manually assigned to subnets and their routers, and DHCP servers had to be configured accordingly. In terms of management, there is a linkage between IP address management and DNS management, because DNS mappings typically need to be appropriately synchronized with IP address assignments. At roughly the same time as DHCP came into widespread use, it became very laborious to manually maintain DNS source files in step with IP address assignments. Because of reverse DNS lookup, it also became necessary to synthesize DNS names even for hosts that only played the role of clients. Therefore, it became necessary to synchronize DHCP server tables with forward and reverse DNS. For this reason, IP address management tools emerged, as discussed for the case of renumbering in . These are, however, centralized solutions that do not exhibit autonomic properties as defined in . A related issue is prefix delegation, especially in IPv6 when more than one prefix may be delegated to the same physical subnet. DHCPv6 Prefix Delegation is a useful solution, but it requires specific configuration so cannot be considered autonomic. How this topic is to be handled in home networks is still in discussion . Still further away is autonomic assignment and delegation of routable IPv4 subnet prefixes. An IPv6 network needs several aspects of host address assignments to be configured. The network might use stateless address autoconfiguration or DHCPv6 in stateless or stateful modes, and there are various alternative forms of Interface Identifier . Another feature is the possibility of Dynamic DNS Update . With appropriate security, this is an automatic approach, where no human intervention is required to create the DNS records for a host after it acquires a new address. However, there are coexistence issues with a traditional DNS setup, as described in .

Since a very early stage, it has been a goal that Internet routing should be self-healing when there is a failure of some kind in the routing system (i.e., a link or a router goes wrong). Also, the problem of finding optimal routes through a network was identified many years ago as a problem in mathematical graph theory, for which well known algorithms were discovered (the Dijkstra and Bellman-Ford algorithms). Thus, routing protocols became largely autonomic from the start, as it was clear that manual configuration of routing tables for a large network was impractical. IGP routers do need some initial configuration data to start up the autonomic routing protocol. Also, BGP-4 routers need detailed static configuration of routing policy data.

Originally, the configuration of a default router in a host was a manual operation. Since the deployment of DHCP, this has been automatic as far as most IPv4 hosts are concerned, but the DHCP server must be appropriately configured. In simple environments such as a home network, the DHCP server resides in the same box as the default router, so this configuration is also automatic. In more complex environments, where an independent DHCP server or a local DHCP relay is used, DHCP configuration is more complex and not automatic. In IPv6 networks, the default router is provided by Router Advertisement messages from the router itself, and all IPv6 hosts make use of it. The router may also provide more complex Route Information Options. The process is essentially autonomic as far as all IPv6 hosts are concerned, and DHCPv6 is not involved. However, there are still open issues when more than one prefix is in use on a subnet, and more than one first-hop router may be available as a result (see, for example, ).

Originally, hostnames were looked up in a static table, often referred to as "hosts.txt" from its traditional file name. When the DNS was deployed during the 1980s, all hosts needed DNS resolver code and needed to be configured with the IP addresses (not the names) of suitable DNS servers. Like the default router, these were originally manually configured. Today, they are provided automatically via DHCP or DHCPv6 . For IPv6 end systems, there is also a way for them to be provided automatically via a Router Advertisement option. However, the DHCP or DHCPv6 server, or the IPv6 router, needs to be configured with the appropriate DNS server addresses. Additionally, some networks deploy Multicast DNS locally to provide additional automation of the name space.

Originally, user authentication and accounting was mainly based on physical connectivity and the degree of trust that follows from direct connectivity. Network operators charged based on the setup of dedicated physical links with users. Automated user authentication was introduced by the Point-to-Point Protocol , and RADIUS protocol in the early 1990s. As long as a user completes online authentication through the RADIUS protocol, the accounting for that user starts on the corresponding Authentication, Authorization, and Accounting (AAA) server automatically. This mechanism enables business models with charging based on the amount of traffic or time. However, user authentication information continues to be manually managed by network administrators. It also becomes complex in the case of mobile users who roam between operators, since prior relationships between the operators are needed.

Security has many aspects that need configuration and are therefore candidates to become autonomic. On the other hand, it is essential that a network's central policy be applied strictly for all security configurations. As a result, security has largely been based on centrally imposed configurations. Many aspects of security depend on policy, for example, password rules, privacy rules, firewall rulesets, intrusion detection and prevention settings, VPN configurations, and the choice of cryptographic algorithms. Policies are, by definition, human made and will therefore also persist in an autonomic environment. However, policies are becoming more high-level, abstracting addressing, for example, and focusing on the user or application. The methods to manage, distribute, and apply policy and to monitor compliance and violations could be autonomic. Today, many security mechanisms show some autonomic properties. For example user authentication via IEEE 802.1x allows automatic mapping of users after authentication into logical contexts (typically VLANs). While today configuration is still very important, the overall mechanism displays signs of self-adaption to changing situations. BGP Flowspec allows a partially autonomic threat-defense mechanism, where threats are identified, the flow information is automatically distributed, and counter-actions can be applied. Today, typically a human operator is still in the loop to check correctness, but over time such mechanisms can become more autonomic. Negotiation capabilities, present in many security protocols, also display simple autonomic behaviors. In this case, a security policy about algorithm strength can be configured into servers but will propagate automatically to clients.

Another area where autonomic processes between peers are involved is state synchronization. In this case, several devices start out with inconsistent state and go through a peer-to-peer procedure after which their states are consistent. Many autonomic or automatic processes include some degree of implicit state synchronization. Network time synchronization is a well-established explicit example, guaranteeing that a participating node's clock state is synchronized with reliable time servers within a defined margin of error, without any overall point of control of the synchronization process.

In current networks, many operations are still heavily dependent on human intelligence and decision, or on centralized top-down network management systems. These operations are the targets of Autonomic Networking technologies. The ultimate goal of Autonomic Networking is to replace human and automated operations by autonomic functions, so that the networks can run independently without depending on a human or Network Management System (NMS) for routine details, while maintaining central control where required. Of course, there would still be the absolute minimum of human input required, particularly during the network-building stage, emergencies, and difficult troubleshooting. This section analyzes the existing human and central dependencies in typical networks and suggests cases where they could, in principle, be replaced by autonomic behaviors.

Building a network requires the operator to analyze the requirements of the new network, design a deployment architecture and topology, decide device locations and capacities, set up hardware, design network services, choose and enable required protocols, configure each device and each protocol, set up central user authentication and accounting policies and databases, design and deploy security mechanisms, etc. Overall, these jobs are quite complex work that cannot become fully autonomic in the foreseeable future. However, part of these jobs may be able to become autonomic, such as detailed device and protocol configurations and database population. The initial network management policies/behaviors may also be transplanted from other networks and automatically localized.

Network maintenance and management are very different for ISP networks and enterprise networks. ISP networks have to change much more frequently than enterprise networks, given the fact that ISP networks have to serve a large number of customers who have very diversified requirements. The current rigid model is that network administrators design a limited number of services for customers to order. New requirements of network services may not be able to be met quickly by human management. Given a real-time request, the response must be autonomic, in order to be flexible and quickly deployed. However, behind the interface, describing abstracted network information and user authorization management may have to depend on human intelligence from network administrators in the foreseeable future. User identification integration/consolidation among networks or network services is another challenge for Autonomic Network access. Currently, many end users have to manually manage their user accounts and authentication information when they switch among networks or network services. Classical network maintenance and management mainly handle the configuration of network devices. Tools have been developed to enable remote management and make such management easier. However, the decision about each configuration detail depends either on human intelligence or rigid templates. Therefore, these are the sources of all network configuration errors -- the human was wrong, the template was wrong, or both were wrong. This is also a barrier to increasing the utility of network resources, because the human managers cannot respond quickly enough to network events, such as traffic bursts, that were not foreseen in the template. For example, currently, a light load is often assumed in network design because there is no mechanism to properly handle a sudden traffic flood. It is therefore common to avoid performance collapses caused by traffic overload by configuring idle resources, with an overprovisioning ratio of at least 2 being normal . There are grounds for concern that the introduction of new, more flexible, methods of network configuration, typified by Software-Defined Networking (SDN), will only make the management problem more complex unless the details are managed automatically or autonomically. There is no doubt that SDN creates both the necessity and the opportunity for automation of configuration management, e.g., . This topic is discussed from a service provider viewpoint in . Autonomic decision processes for configuration would enable dynamic management of network resources (by managing resource-relevant configuration). Self-adapting network configuration would adjust the network into the best possible situation; this would prevent configuration errors from having lasting impact.

Setting up security for a network generally requires very detailed human intervention or relies entirely on default configurations that may be too strict or too risky for the particular situation of the network. While some aspects of security are intrinsically top-down in nature (e.g., broadcasting a specific security policy to all hosts), others could be self-managed within the network. In an Autonomic Network, where nodes within a domain have a mutually verifiable domain identity, security processes could run entirely automatically. Nodes could identify each other securely, negotiating required security settings and even shared keys if needed. The locations of the trust anchors (certificate authority, registration authority), certificate revocation lists, policy server, etc., can be found by service discovery. Transactions such as a download of a certificate revocation list can be authenticated via a common trust anchor. Policy distribution can also be entirely automated and secured via a common trust anchor. These concepts lead to a network where the intrinsic security is automatic and applied by default, i.e., a "self-protecting" network. For further discussion, see .

Current networks suffer difficulties in locating the cause of network failures. Although network devices may issue many warnings while running, most of them are not sufficiently precise to be identified as errors. Some of them are early warnings that would not develop into real errors. Others are, in effect, random noise. During a major failure, many different devices will issue multiple warnings within a short time, causing overload for the NMS and the operators. However, for many scenarios, human experience is still vital to identify real issues and locate them. This situation may be improved by automatically associating warnings from multiple network devices together. Also, introducing automated learning techniques (comparing current warnings with historical relationships between warnings and actual faults) could increase the possibility and success rate of Autonomic Network diagnoses and troubleshooting. Depending on the network errors, some of them (particularly hardware failures) may always require human intervention. However, Autonomic Network management behavior may help to reduce the impact of errors, for example, by switching traffic flows around. Today, this is usually manual (except for classical routing updates). Fixing software failures and configuration errors currently depends on humans and may even involve rolling back software versions and rebooting hardware. Such problems could be autonomically corrected if there were diagnostics and recovery functions defined in advance for them. This would fulfill the concept of self-healing. Another possible autonomic function is predicting device failures or overloads before they occur. A device could predict its own failure and warn its neighbors, or a device could predict its neighbor's failure. In either case, an Autonomic Network could respond as if the failure had already occurred by routing around the problem and reporting the failure, with no disturbance to users. The criteria for predicting failure could be temperature, battery status, bit error rates, etc. The criteria for predicting overload could be increasing load factor, latency, jitter, congestion loss, etc.

There are innumerable properties of network devices and end systems that today need to be configured either manually, by scripting, or by using a management protocol such as the Network Configuration Protocol (NETCONF) . In an Autonomic Network, all of these would need to either have satisfactory default values or be configured automatically. Some examples are parameters for tunnels of various kinds, flows (in an SDN context), quality of service, service function chaining, energy management, system identification, and NTP configuration, but the list is endless. The task of Autonomic Networking is to incrementally build up individual autonomic processes that could progressively be combined to respond to every type of network event. Building on the preceding background information, and on the reference model in , this section outlines the gaps and missing features in general terms and, in some cases, mentions general design principles that should apply.

Network services are dependent on a number of devices and parameters to be in place in a certain order. For example, after a power failure, a coordinated sequence of "return to normal" operations is desirable (e.g., switches and routers first, DNS servers second, etc.). Today, the correct sequence of events is either known only by a human administrator or automated in a central script. In a truly Autonomic Network, elements should understand their dependencies and be able to resolve them locally. In order to make right or good decisions autonomically, the network devices need to know more information than just reachability (routing) information from the relevant or neighbor devices. Devices must be able to derive, for themselves, the dependencies between such information and configurations. There are therefore increased requirements for horizontal information exchange in the networks. Particularly, three types of interaction among peer network devices are needed for autonomic decisions: discovery (to find neighbors and peers), synchronization (to agree on network status), and negotiation (when things need to be changed). Thus, there is a need for reusable discovery, synchronization, and negotiation mechanisms that would support the discovery of many different types of device, the synchronization of many types of parameter, and the negotiation of many different types of objective.

Elements of autonomic functions already exist today, within many different protocols. However, all such functions have their own discovery, transport, messaging, and security mechanisms as well as non-autonomic management interfaces. Each protocol has its own version of the above-mentioned functions to serve specific and narrow purposes. It is often difficult to extend an existing protocol to serve different purposes. Therefore, in order to provide the reusable discovery, synchronization, and negotiation mechanisms mentioned above, it is desirable to develop a set of reusable common protocol components for Autonomic Networking. These components should be: Able to identify other devices, users, and processes securely. Able to automatically secure operations, based on the above identity scheme. Able to manage any type of information and information flows. Able to discover peer devices and services for various Autonomic Service Agents (or autonomic functions). Able to support closed-loop operations when needed to provide self-managing functions involving more than one device. Separable from the specific Autonomic Service Agents (or autonomic functions). Reusable by other autonomic functions.

The common components will, in effect, act as a control plane for autonomic operations. This control plane might be implemented in-band as functions of the target network, in an overlay network, or even out-of-band in a separate network. Autonomic operations will be capable of changing how the network operates and allocating resources without human intervention or knowledge, so it is essential that they are secure. Therefore, the control plane must be designed to be secure against forged autonomic operations and man-in-the middle attacks and as secure as reasonably possible against denial-of-service attacks. It must be decided whether the control plane needs to be resistant to unwanted monitoring, i.e., whether encryption is required.

Many existing protocols have been defined to be as flexible as possible. Consequently, these protocols need numerous initial configurations to start operations. There are choices and options that are irrelevant in any particular case, some of which target corner cases. Furthermore, in protocols that have existed for years, some design considerations are no longer relevant, since the underlying hardware technologies have evolved meanwhile. To appreciate the scale of this problem, consider that more than 160 DHCP options have been defined for IPv4. Even sample router configuration files readily available online contain more than 200 lines of commands. There is therefore considerable scope for simplifying the operational tools for configuration of common protocols, even if the underlying protocols themselves cannot be simplified. From another perspective, the deep reason why human decisions are often needed mainly results from the lack of information. When a device can collect enough information horizontally from other devices, it should be able to decide many parameters by itself, instead of receiving them from top-down configuration. It is desired that top-down management is reduced in Autonomic Networking. Ideally, only the abstract Intent is needed from the human administrators. Neither users nor administrators should need to create and maintain detailed policies and profiles; if they are needed, they should be built autonomically. The local parameters should be decided by distributed Autonomic Nodes themselves, either from historic knowledge, analytics of current conditions, closed logical decision loops, or a combination of all.

In a conventional network, there is no mechanism for trying something out safely, which means that configuration changes have to be designed in the abstract and their probable effects have to be estimated theoretically. In principle, an alternative to this would be to test the changes on a complete and realistic network simulator. However, this is a practical impossibility for a large network that is constantly changing, even if an accurate simulation could be performed. There is therefore a risk that applying changes to a running network will cause a failure of some kind. An autonomic network could fill this gap by supporting a closed loop "dry run" mode in which each configuration change could be tested out dynamically in the control plane without actually affecting the data plane. If the results are satisfactory, the change could be made live on the running network. If there is a consistency problem such as overcommitment of resources or incompatibility with another configuration setting, the change could be rolled back dynamically with no impact on traffic or users.

The more knowledge and experience we have, the better decisions we can make. It is the same for networks and network management. When one component in the network lacks knowledge that affects what it should do, and another component has that knowledge, we usually rely on a human operator or a centralized management tool to convey the knowledge. Up to now, the only available network knowledge is usually the current network status inside a given device or relevant current status from other devices. However, historic knowledge is very helpful to make correct decisions, in particular, to reduce network oscillation or to manage network resources over time. Transplantable knowledge from other networks can be helpful to initially set up a new network or new network devices. Knowledge of relationships between network events and network configuration may help a network to decide the best parameters according to real performance feedback. In addition to such historic knowledge, powerful data analytics of current network conditions may also be a valuable source of knowledge that can be exploited directly by Autonomic Nodes.

This document is focused on what is missing to allow autonomic network configuration, including security settings, of course. Therefore, it does not itself create any new security issues. It is worth underlining that autonomic technology must be designed with strong security properties from the start, since a network with vulnerable autonomic functions would be at great risk.

The authors would like to acknowledge the valuable comments made by participants in the IRTF Network Management Research Group. Reviews by Kevin Fall and Rene Struik were especially helpful.