Beneficial Functions of Middleboxes

From RFC3234, "A middlebox is defined as any intermediary device performing functions other than the normal, standard functions of an IP router on the datagram path between a source host and destination host." Middleboxes are usually (but not exclusively) deployed at locations permitting observation of bidirectional traffic flows. Such locations are typically points where stub networks connect to the Internet; e.g.,: Where a residential or business customer connects to its service provider(s), which may include multi-homing. On the Gi interface where a GGSN connects to a PDN (see section 3.1 of ). The QUIC working group and PLUS BoF are debating the appropriate amount of information that end-points should expose to on-path network middleboxes and human trouble-shooters. (Some information used for debugging is discussed in .) This document itemizes a variety of features provided by middleboxes and by ad hoc analysis performed by operators using packet analyzers. Many of the techniques described in this document require stateful analysis of transport streams. A generic state machine is described in . Although many middleboxes observe and manipulate application-layer content (e.g., session boarder controllers ) they are out of scope for this document, the aim being to describe benefits of middleboxes using transport-layer features. An earlier document describes the impact of pervasive encryption of application-layer data on network monitoring, protecting and troubleshooting. This document advocates for transport connections to be measured and managed by the network for the benefit of both parties: for the end-user to receive better quality of experience, and for the network operator to improve resource usage, the former being a consequence of the latter. This document does not discuss whether exposing some data to on-path devices for network assistance purposes can be achieved by using in-band or out-of-band mechanisms.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

A number of measurements can be made by network devices that are either in-line with the traffic (responsible for forwarding) or receiving off-line copy of traffic from a tap or file capture. These measurements can be used either by automated systems, or for manual network troubleshooting purposes (e.g., using packet analysis tools). The automated systems can further be classified as monitoring systems that compute performance indicators for large numbers of connections and generate aggregated reports from them, and active systems that make decisions on how to handle specific packets based on these performance indicators. Long-term trends in these measurements can aid an operator in capacity planning. Short-term anomalies revealed by these measurements can identify network breakages, attacks in progress, or misbehaving devices/applications.

Network problems and under-provisioning can be detected if packet loss is measurable. TCP packet loss can be detected by observing gaps in sequence numbers, retransmitted sequence numbers, and SACK options. Packet loss can be detected per direction. Gaps indicate loss upstream of the tap point; retransmissions indicate loss downstream of the tap. Selective acknowledgements (SACKs) can be used to detect either upstream or downstream packet loss (although some care needs to be taken to avoid mis-identifying packet reordering as packet loss), and to distinguish between upstream vs. downstream losses. Packet loss measurements on both sides of the measurement point are an important component in precisely diagnosing insufficiently dimensioned devices or links in networks. Additionally, since packet losses are one of the two main ways for congestion to manifest (the other being queueing delay), packet loss is an important measurement for any middlebox that needs to make traffic handling decisions based on observed levels of congestion.

A TCP packet stream can be used to measure the round-trip time on each side of the measurement point. During the connection handshake, the SYN, SYNACK, and ACK timings can be used to establish a baseline RTT in each direction. Once the connection is established, the RTT between the server and the measurement point can only reliably be determined using TCP timestamps. On the side between the measurement point and the client, the exact timing of data segments and ACKs can be used as an alternative. For this latter method to be accurate when packet loss is present, the connection must use selective acknowledgements. In many networks, congestion will show up as increasing packet queueing, and congestion-induced packet loss will only happen in extreme cases. RTTs will also show up as a much smoother signal than the discrete packet loss events. This makes RTTs a good way to identify individual subscribers for whom the network is a bottleneck at a given time, or geographical sites (such as cellular towers) that are experiencing large scale congestion. The main limit of RTT measurement as a congestion signal is the difficulty of reliably distinguishing between the data segments being queued vs. the ACKs being queued.

If a network is reordering packets of transport connections, caused perhaps by ECMP misconfiguration (e.g., described in and ), the end-points may react as if packet loss is occurring and retransmit packets or reduce forwarding rates. It is therefore beneficial to be able to diagnose packet reordering from within a network. For TCP, packet reordering can be detected by observing TCP sequence numbers per direction. See for example a number of standard packet reordering metrics in and informational metrics in .

Although throughput to or from an IP address can be measured without transport-layer measurements, the transport layer provides clues about what the end-points were attempting to do. One way of quickly excluding the network as the bottleneck during troubleshooting is to check whether the speed is limited by the endpoints. For example, the connection speed might instead be limited by suboptimal TCP options, the sender's congestion window, the sender temporarily running out of data to send, the sender waiting for the receiver to send another request, or the receiver closing the receive window. This data is also useful for middleboxes used to measure network quality of service. Connections, or portions of connections, that are limited by the endpoints do not provide an accurate measure of network's speed, and can be discounted or completely excluded in such analyses.

When an application or network resource is under attack, it is useful to identify this situation from the network perspective, upstream of the attacked resource. Although detection methods tend to be proprietary, DDoS attack detection is fundamentally one of: detecting protocol violations by tracking the transport-layer state machine or application-layer messaging; or anomaly detection by noticing atypical traffic patterns taken from measurements. Two trends in protocol design will make DDoS detection more difficult: the desire to encrypt transport-layer communication and sequence numbers; the desire to avoid statistical fingerprinting by adding entropy in various forms. Those desires assist in the worthy goal of improved privacy, but also serve to defeat DDoS detection.

One notable source of packet loss is packet corruption. This corruption will generally not be detected until the checksums are validated by the endpoint, and the packet is dropped. This means that detecting the exact location where packets are lost is not sufficient when troubleshooting networks. It should also be possible to find out where packets are being corrupted. IP and TCP checksum verification allows a measurement device to correctly distinguish between upstream packet corruption and normal downstream packet loss. Transport protocol designers should consider whether a middlebox will be able to detect corrupted or tampered packets.

Network health may also be gleaned from application-layer diagnosis. E.g., DNS response times and retransmissions by correlating answers to queries. Various protocol-aware voice and video quality analysis. Could this type of information be provided in a transport layer?

This section describes features provided by in-line devices that go beyond measurement by modifying, discarding, delaying, or prioritizing traffic.

Network Address Translators (NATs) allow multiple devices to share a public address by dividing the transport-layer port space among the devices. NAT behavior recommendations are found for UDP in BCP 127 and for TCP in BCP 142 . To support NAT, there must be transport-layer port numbers that can be modified by the network. The application-layer must not assume the port number was left unchanged (e.g., by including it in a checksum or signing it). Address sharing is also used in the context of IPv6 transition. For example, DS-Lite AFTR, NAT64, or MAP-* are features that are enabled in the network to allow for IPv4 service continuity over an IPv6 network. Further, because of some multi-homing considerations, IPv6 prefix translation may be enabled by some enterprises by means of NPTv6.

Firewalls are pervasive and essential components that inspect incoming and outgoing traffic. Firewalls are usually the cornerstone of a security policy that is enforced in end-user premises and other locations to provide strict guarantees about traffic that may be authorized to enter/leave the said premises, as well as end-users who may be assigned different clearance levels regarding which networks and portions of the Internet they may acess. Arguably many users within various types of organizations would not have been granted Internet access if not for safety provided by firewalls. An important aspect of a firewall policy is differentiating internally-initiated from externally-initiated communications. For TCP, this is easily done by tracking the TCP state machine. Furthermore, the ending of a TCP connection is indicated by RST or FIN flags. For UDP, the firewall can be opened if the first packet comes from an internal user, but the closing is generally done by an idle timer of arbitrary duration, which might not match the expectations of the application. Simple IPv6 firewall capabilities for customer premises equipment (both stateless and stateful) are described in . A firewall functions better when it can observe the protocol state machine, described generally by Transport-Independent Path Layer State Management.

In the context of a distributed denial-of-service (DDoS) attack, the purpose of a scrubber is to discard attack traffic while permitting useful traffic. E.g., such a mitigator is described in . When attacks occur against constrained resources, there is obviously a huge benefit in being able to scrub well. Furthermore, this is solely a task for an on-path network device because neither end-point of a legitimate connection has any control over the source of the attack traffic. Source-spoofed DDoS attacks can be mitigated at the source using BCP 38 (), but it is more difficult if source address filtering cannot be applied. In contrast to devices in the core of the Internet, middleboxes statefully observing bidirectional transport connections can reject source-spoofed TCP traffic based on the inability to provide sensible acknowledgement numbers to complete the three-way handshake. Obviously this requires middlebox visibility into transport-layer state machine. Middleboxes may also scrub on the basis of statistical classification: testing how likely a given packet is legitimate. As protocol designers add more entropy to headers and lengths, this test becomes less useful and the best scrubbing strategy becomes random drop.

In order to enhance the end-user's quality of experience, some operators deploy implicit identification features that rely upon the correlation of network-related information to access some local services. For example, service portals operated by some operators may be accessed immediately by end-users without any explicit identification for the sake of improved service availability. This is doable thanks to on-path devices that inject appropriate metadata that can be used by the remote server to enforce per-subscriber policies. The information can be injected at the application layer or at the transport layer (when an address sharing mechanism is in use). An experimental implementation using a TCP option is described in . For the intended use of implicit identification, it is more secure to have a trusted middlebox mark this traffic than to trust end-user devices.

Performance-Enhancing Proxies (PEPs) can improve performance in some types of networks by improving packet spacing or generating local acknowledgements, and are most commonly used in satellite and cellular networks. Transport-Layer PEPs are described in section 2.1.1 of . PEPs allow central deployment of congestion control algorithms more suited to the specific network, most commonly use of delay-based congestion control. More advanced TCP PEPs deploy congestion control systems that treat all of a single end-user's TCP connections as a single unit, improving fairness and allowing faster reaction to changing network conditions. Local acknowledgements generated by PEPs speed up TCP slow start by splitting the effective latency, and allow for retransmissions to be done from the PEP rather than from the actual sender, saving downlink bandwidth on retransmissions. Local acknowledgements will also allow a PEP to maintain a local buffer of data appropriate to the actual network conditions, whereas the actual endpoints would often send too much or too little. A PEP function requires transport-layer fields that allow chunks of data to be identified (e.g., TCP sequence numbers), acknowledgements to be identified (e.g., TCP ACK numbers), and acknowledgements to be created from the PEP. Note that PEPs are only useful in some types of networks, and poor design could make performance worse.

Network Coding is a technique for compressing traffic or adding redundancy for transmission over low-bandwidth, long-latency links such as satellite links. One method is to deploy network-coding gateways at each end of those links, with a network-coding tunnel between them via the slow/lossy/long-latency links. The network coding gateways may employ some techniques of PEPs, such as creating acknowledgements of queued data, removing retransmissions and pacing data rates to reduce queue oscillation.

The Hybrid Access Aggregation Point (HAAP) is a middlebox that allows customers to aggregate the bandwidth of multiple access technologies . One of the approaches uses MPTCP proxies to forward traffic along multiple paths. The MPTCP proxy operates at the transport layer while being located in the operator's network. The support of multipath transport capabilities by communicating hosts remains a privileged target design so that such hosts can directly use the available resources provided by a variety of access networks they can connect to. Nevertheless, network operators do not control end hosts while the support of MPTCP by content servers remains marginal. Network-Assisted MPTCP deployment models are designed to facilitate the adoption of MPTCP for the establishment of multi-path communications without making any assumption about the support of MPTCP capabilities by communicating peers. Network-Assisted MPTCP deployment models rely upon MPTCP Conversion Points (MCPs) that act on behalf of hosts so that they can take advantage of establishing communications over multiple paths . Note that an MPTCP proxy can be beneficial even if both the client and the server are MPTCP-compliant. Examples of such cases are listed below: The use of private IPv4 addresses in some access networks. Typically, additional subflows can not be added to the MPTCP connection without the help of an MCP. The assignment of IPv6 prefixes only by some networks. If the server is IPv4-only, IPv6 subflows cannot be added to an MPTCP connection established with that server, by definition. Subscription to some service offerings is subject to volume quota.

Bulk traffic may be served with a higher latency than interactive traffic with no reduction in throughput. This fact allows a middlebox function to improve response times in interactive applications by prioritizing, policing, or remarking interactive transport connections differently from bulk traffic transport connections. E.g., gaming traffic may be prioritized over email or software updates. Middleboxes may identify different classes of traffic by inspecting multiple layers of header and payload.

Basic traffic shaping functionality requires no transport-layer information. All that is needed is a way of mapping each packet to a traffic shaper quota. For example, there may be a rate limit per 5-tuple or per subscriber IP address. However, such fixed traffic shaping rules are wasteful as they end up rate limiting traffic even when the network has free resources available. More advanced traffic shaping devices use transport layer metrics described in to detect congestion on either a per-site or per-user level, and use different traffic shaping rules when congestion is detected. This type of device can overcome limitations of down-stream devices that behave poorly (e.g., by excessive buffering or sub-optimally dropping packets).

Several service offerings rely upon a volume-based charging model. Operators may assist end-users in conserving their data quota by deploying on-path functions that shape traffic that would otherwise be agressively transferred. For example, a fast download of a video that won't be viewed completely by the subscriber may lead to quick exhaustion of the data quota. Limiting the video download rate conserves quota for the benefit of the end-user.

The authors thank Brian Trammell and Brian Carpenter for their review and suggestions.

This memo includes no request to IANA.

This document intentionally excludes middleboxes that observe or manipulate application-layer data. The benefits described in this document can all be implemented without violating confidentiality. However, there is always the question of whether the fields and packet properties used to achieve these benefits may also be used for harm. In particular, we want to ask what confidentiality is lost by exposing transport-layer fields beyond what can be learned by observing IP-layer fields. Sequence numbers: an observer can learn how much data is transferred. Start/Stop indicators: an observer can count transactions for some applications. Device fingerprinting: an observer may be more easily able to identify a device type when different devices use different default field values or options.

Being able to observe sequence numbers or session identifiers may make it easier to modify or terminate a transport connection. E.g., observing TCP sequence numbers allows generation of a RST packet that terminates the connection. However, signing transport fields mitigates this attack. The attack and solution are described for the TCP authentication option .

Proposition: network maintainability and security can be improved by providing firewalls and DDoS mechanisms with some information about transport connections. In contrast, it would be very difficult to secure a network in which every packet appears unique and filled with random bits. For denial-of-service (DoS) attacks on bandwidth, the receiving end-point is usually on the wrong side of the constrained network link. This fact makes it seem reasonable to give some clues to allow a middlebox device to help out before the constrained link. E.g., in a blind attack, an attacker cannot receive data from the target of the attack (section 4.6.3.2 of ). In the case of TCP, the blind attacker cannot complete the three-way handshake. In the balance, some features providing the ability to mitigate/filter attacks and fix broken networks will improve security vs. the scenario when all packets are completely opaque.