]> University of Aberdeen

Department of Engineering Fraser Noble Building Aberdeen AB24 3UE Scotland gorry@erg.abdn.ac.uk http://www.erg.abdn.ac.uk/

Transport TSVWG transport design, operations and management This document describes the implications of applying end-to-end encryption at the transport layer. It identifies some in-network uses of transport layer header information that can be used with a transport header integrity check. It reviews the implication of developing encrypted end-to-end transport protocols and examines the implication of developing and deploying encrypted end-to-end transport protocols.

This document discusses the implications of end-to-end encryption applied at the transport layer, and examines the impact on transport protocol design, transport use, and network operations and management. It also considers some anticipated implications on transport and application evolution. The transport layer is the first end-to-end layer in the network stack. Despite headers having end-to-end meaning, some transport headers have come to be used in various ways within the Internet. In response to pervasive monitoring revelations and the IETF consensus that "Pervasive Monitoring is an Attack" , efforts are underway to increase encryption of Internet traffic, which would prevent visibility of transport headers. This has implications on how network protocols are designed and used . Transport information that is sent without end-to-end integrity check could be modified by "middleboxes" - defined as any intermediary box performing functions apart from normal, standard functions of an IP router on the data path between a source host and destination host . When transport headers are modified by network devices on the path, this can change the end-to-end protocol transport protocol behaviour in a way that may have benefits (e.g., to user performance/cost) or may hinder (e.g., disrupting application experience). Whatever the outcome, modification of packets by a middlebox was not usually intended when the protocol was specified and is usually not known by the sending or receiving endpoints. Middleboxes have been deployed for a variety of reasons , including protocol enhancement, proxies such as Protocol Enhancing Proxies (PEPs) , TCP acknowledgement (ACK) enhancement , use by application protocol caches , application layer gateways , etc. summarizes some of the functions provided by such middleboxes, and benefits that may arise when used in specific deployment scenarios. Such methods, which involve in-network modification of transport headers, are not further discussed. Transport protocols can be designed to encrypt or authenticate transport header fields. Authentication methods at the transport layer can detect any changes to an immutable header field that were made by a network device along a path. These methods do not require encryption of the header fields, and hence authenticated fields may remain visible to network devices. A receiving transport endpoint can use an integrity check to avoid accepting modified protocol headers. This document therefore considers the case where transport header fields are not altered as a packet traverses the network path. Authentication methods (that provide integrity checks of protocols fields) have also been specified at the network layer, and this also protects transport header fields. The network layer itself carries protocol header fields that are increasingly used to help forwarding decisions reflect the need of transport protocols, such the IPv6 Flow Label , the Differentiated Services Code Point (DSCP) and Explicit Congestion Notification (ECN) . Encryption methods can hide information from an eavesdropper in the network. Encryption can also help protect the privacy of a user, by hiding data relating to user/device identity or location. Neither an integrity check nor encryption methods prevent traffic analysis, and usage needs to reflect that profiling of users and fingerprinting of behaviour can take place even on encrypted traffic flows. This document seeks to identify the implications of various approaches to transport protocol authentication and encryption.

End-to-end encryption can be applied at various protocol layers. It can be applied above the transport to encrypt the transport payload. One motive to use encryption is a response to perceptions that the network has become ossified by over-reliance on middleboxes that prevent new protocols and mechanisms from being deployed. This has lead to a common perception that there is too much “manipulation” of protocol headers within the network, and that designing to deploy in such networks is preventing transport evolution. In the light of this, a method that authenticates transport headers may help improve the pace of transport development, by eliminating the need to always consider deployed middleboxes , or potentially to only explicitly enable middlebox use for particular paths with particular middleboxes . Another perspective stems from increased concerns about privacy and surveillance. Some Internet users have valued the ability to protect identity and defend against traffic analysis, and have used methods such as IPsec ESP and Tor . Revelations about the use of pervasive surveillance have, to some extent, eroded trust in the service offered by network operators, and following the Snowden revelation in the USA in 2013 has led to an increased desire for people to employ encryption to avoid unwanted "eavesdropping" on their communications. Whatever the reasons, there are now activities in the IETF to design new protocols that may include some form of transport header encryption (e.g., QUIC ). The use of transport layer authentication and encryption exposes a tussle between middlebox vendors, operators, applications developers and users. On the one hand, future Internet protocols that enable large-scale encryption assist in the restoration of the end-to-end nature of the Internet by returning complex processing to the endpoints, since middleboxes cannot modify what they cannot see. On the other hand, encryption of transport layer header information has implications to people responsible for operating networks and researchers and analysts seeking to understand the dynamics of protocols and traffic patterns. Whatever the motives, a decision to use pervasive of transport header encryption will have implications on the way in which design and evaluation is performed, and which can in turn impact the direction of evolution of the TCP/IP stack. The next subsections briefly review some security design options for transport protocols.

Transport layer header information can be authenticated. An integrity check that protects the immutable transport header fields, but can still expose the transport protocol header information in the clear, allowing in-network devices to observes these fields. An integrity check can not prevent in-network modification, but can avoid accepting changes and avoid impact on the transport protocol operation. An example transport authentication mechanism is TCP-Authentication (TCP-AO) . This TCP option authenticates TCP segments, including the IP pseudo header, TCP header, and TCP data. TCP-AO protects the transport layer, preventing attacks from disabling the TCP connection itself. TCP-AO may interact with middleboxes, depending on their behavior . The IPSec Authentication Header (AH) works at the network layer and authenticates the IP payload. This therefore also authenticates all transport headers, and verifies their integrity at the receiver, preventing in-network modification.

The transport layer payload can be encrypted to protect the content of transport segments. This leaves transport protocol header information in the clear. The integrity of immutable transport header fields could be protected by combining this with an integrity check (). Examples of encrypting the payload include Transport Layer Security (TLS) over TCP or Datagram TLS (DTLS) over UDP .

The network layer payload could be encrypted (including the entire transport header and payload). This method does not expose any transport information to devices in the network, which also prevents modification along the network path. The IPSec Encapsulating Security Payload (ESP) is an example of encryption at the network layer, it encrypts and authenticates all transport headers, preventing visibility of the headers by in-network devices. Some Virtual Private Network (VPN) methods also encrypt these headers.

A transport protocol design can encrypt selected header fields, while also choosing to authenticate fields in the transport header. This allows specific transport header fields to be made observable by network devices. End-to end integrity checks can prevent an endpoint from undetected modification of the immutable transport headers. The choice of which fields to expose and which to encrypt is a design choice for the transport protocol. Any selective encryption method requires trading two conflicting goals for a transport protocol designer to decide which header fields to encrypt. On the one hand, security work typically employs a design technique that seeks to expose only what is needed. On the other hand, there may be performance and operational benefits in exposing selected information to network tools. Mutable fields in the transport header provide opportunities for middleboxes to modify the transport behaviour (e.g., the extended headers described in ). This considers only immutable fields in the transport headers, that is, fields that may be authenticated end-to-end across a path. An example of a method that encrypts some, but not all, transport information is UDP-in-GRE when used with GRE encryption.

The transport information can be made visible in a network-layer header. This has the advantage that this information can then be observed by in-network devices. This has the advantage that a single header can support all transport protocols, but there may also be less desirable implications of separating the operation of the transport protocol from the measurement framework. Some measurements may be made by adding additional protocol headers carrying operations, administration and management (OAM) information to packets at the ingress to a maintenance domain (e.g., an Ethernet protocol header with timestamps and sequence number information using a method such as 802.11ag) and removing the additional header at the egress of the maintenance domain. This approach enables some types of measurements, but does not cover the entire range of measurements described in this document. Another example of a network-layer approach is the IPv6 Performance and Diagnostic Metrics (PDM) Destination Option . This allows a sender to optionally include a destination option that caries header fields that can be used to observe timestamps and packet sequence numbers. This information could be authenticated by receiving transport endpoints when the information is added at the sender and visible at the receiving endpoint, although methods to do this have not currently been proposed. This method needs to be explicitly enabled at the sender. A drawback of using extension headers is that IPv4 network options are often not supported (or are carried on a slower processing path) and some IPv6 networks are also known to drop packets that set an IPv6 header extension. Another disadvantage is that protocols that separately expose header information do not necessarily have an advantage to expose the information that is utilised by the protocol itself, and could manipulate this header information to gain an advantage from the network.

This section identifies ways that actors can benefit by observing (non-encrypted) transport header fields at devices in the network. The list of actors who perform measurements include: Protocol developers and implementors of TCP/IP stacks; Researchers working on new mechanisms; Use of new applications using existing applications; Analysis researching the impact of mechanisms on network equipment or specific network topologies; Staff supporting operation of a network. One approach is to use active measurement using dedicated tools to generate and measure test traffic. To test a transport path, such active tools need to be run from an endpoint, and most operators do not have access to user equipment. There may also be costs associated with running such tests (e.g., the implications of bandwidth tests in a mobile network are obvious). Some active measurements (e.g., response under load or particular workloads) may perturb other traffic, and could require dedicated access to the network segment. An alternative approach is to use in-network techniques that observe transport packet headers in operational networks to make the measurements. Transport layer information can help identify whether the link/network tuning is effective and alert to potential problems that can be hard to derive from link or device measurements alone. The design trade offs for radio networks are often very different to those of wired networks. A radio-based network (e.g., cellular mobile, enterprise WiFi, satellite access/backhaul, point-to-point radio) has the complexity of a subsystem that performs radio resource management - with direct impact on the available capacity, and potentially loss/reordering of packets. The impact of the pattern of loss and congestion, differs for different traffic types, correlation with propagation and interference can all have significant impact on the cost and performance of a provided service. The need for this type of information is expected to increase as operators bring together heterogeneous types of network equipment and seek to deploy opportunistic methods to access radio spectrum. In-network observation of transport protocol headers requires knowledge of the format of the transport header: Flows, need to be identified at the level required for monitoring; The protocol and version of the header that is being used. As protocols evolve over time and there may be a need to introduce new transport headers. This may require interpretation of protocol version information or connection setup information; The position and syntax of any transport headers that need to be observed. IETF transport protocols specify this information. The following subsections describe various ways that observable transport information may be utilised.

Transport protocol header information can identify a flow and the connection state of the flow, together with the protocol options being used. In some usages, a low-numbered (well-known ) port that can identify a protocol (although port information alone is not sufficient to guarantee identification of a protocol). Transport protocols, such as TCP and SCTP specify a standard base header that includes sequence number information and other data, with the possibility to negotiate additional headers at connection setup and identified by an option number in the transport header. UDP-based protocols sometimes do not use well-known ports but also can instead be identified by signalling protocols or through the use of magic numbers placed in the first byte(s) of the datagram payload.

Some actors have a need to characterise the performance of link/network segments. Passive monitoring uses observed traffic to makes inferences from transport headers to derive measurements. A variety of open source and commercial tools can utilise this information. Transport fields in the Real Time Protocol (RTP) header can be observed to derive traffic volume measurements and provide information on the progress and quality of a session using RTP. Key performance indicators are retransmission rate, packet drop rate, sector utilization level, a measure of reordering, peak rate, the CE-marking rate, etc. Metadata is often important to understand the context under which the data was collected, including the time, observation point, and way in which metrics were accumulated. Some Internet transports report summary performance data that is observable in the network (e.g., RTCP feedback). A user of summary measurement data needs to trust the source of this data and the method used to generate the summary information. When encryption conceals information in packet headers, measurements need to rely on pattern inferences and other heuristics grows, and accuracy suffers .

Transport headers may be observed to derive volume measures per-application, to characterise the traffic using a network segment and pattern of network usage. This may be measured per endpoint or aggregate of endpoint (e.g., by an operator to assess subscriber usage). This can also be used to trigger measurement-based traffic shaping and to implement QoS support within the network and lower layers. Volume measures can be valuable for capacity planning (providing detail of trends rather than the volume per subscriber).

Flow loss rate is often used as a metric for performance assessment and to characterise the transport behaviour. Understanding the root cause of loss can help an operator determine whether this requires corrective action. There are various cause of loss, including: corruption on a link (e.g., interference on a radio link), buffer overflow (e.g., due to congestion), policing (traffic management), buffer management (e.g., Active Queue Management (AQM). Loss can be monitored at the interface level by devices in the network. It is often important to understand the conditions under which packet loss occurs, which usually means relating loss to the traffic flowing on the network segment at the time of loss. Understanding flow loss rate requires either maintaining per flow packet counters or by observing sequence numbers in transport headers. Observation of transport feedback information (observing loss reports, e.g., RTCP, TCP SACK) can increase understanding of the impact of loss and help identify cases where loss may have been wrongly identified, or the transport did not require the lost packet. It is sometimes more important to understand the pattern of loss, than the loss rate - since losses can often occur as bursts, rather than randomly timed events.

The throughput observed by a flow can be determined even when a flow is encrypted, providing the individual flow can be identified. Goodput is a measure of useful data exchanged (the ratio of useful/total volume of traffic sent by a flow), which requires ability to differentiate loss and retransmission of packets (e.g., by observing packet sequence numbers in TCP or RTP).

Latency is a key performance metric that impacts application response time and user-perceived response time. It also often indirectly impacts throughput and flow completion time. Latency determines the reaction time of the transport protocol itself, impacting flow setup, congestion control, loss recovery, and other transport mechanisms. The observed latency can have many components . Of these, unnecessary/unwanted queuing in network buffers has often been observed as a significant factor. Once the cause of unwanted latency has been identified, this can often be eliminated, and determining latency metrics is a key driver in the deployment of AQM , DiffServ , and ECN . To measure latency across a part of the path, an observation point can measure the experienced round trip time (RTT) using packet sequence numbers, and acknowledgements, or by observing header timestamp information. Such information allows an observation point in the network to determine not only the path RTT, but also to measure the upstream and downstream contribution to the RTT. This may be used to locate a source of latency, e.g., by observing cases where the ratio of median to minimum RTT is large for a part of a path. An example usage of this method could be to identify excessive buffers and to deploy or configure Active Queue Management (AQM) . Operators deploying such tools can effectively eliminate unnecessary queuing in routers and other devices. AQM methods need to be deployed at the capacity bottleneck, but are often deployed in combination with other techniques, such as scheduling and although parameter-less methods are desired , current methods often cannot scale across all possible deployment scenarios. The service offered by operators can therefore benefit from latency information to understand the impact of deployment and tune deployed services. Some network applications are sensitive to packet jitter, and it can be necessary to measure the jitter observed along a portion of the path. The requirements to measure jitter resemble those for the measurement of latency.

Significant flow reordering can impact time-critical applications and can be interpreted as loss by reliable transports. Many transport protocol techniques are impacted by reordering (e.g., triggering TCP retransmission, or rebuffering of real-time applications). Packet reordering can occur for many reasons (from equipment design to misconfiguration of forwarding rules). As in the drive to reduce network latency, there is a need for operational tools to be able to detect misordered packet flows and quantify the degree or reordering. Techniques for measuring reordering typically observe packet sequence numbers. Metrics have been defined that evaluate whether a network has maintained packet order on a packet-by-packet basis and . There has been initiatives in the IETF transport area to reduce the impact of reordering withing a transport flow, possibly leading to reduced the requirements for ordering. These have promise to simplify network equipment design as well as the potential to improve robustness of the transport service. Measurements of reordering can help understand the level of reordering within deployed infrastructure, and inform decisions about how to progress such mechanisms.

This section describes transport information that is already observable in network-layer header fields.

Endpoints should expose flow information in the IPv6 Flow Label field of the network-layer header (e..g. ). This can be used to inform network-layer queuing, forwarding (e.g., for equal cost multi-path (ECMP) routing, and Link Aggregation (LAG)). This can provide useful information to assign packets to flows in the data collected by measurement campaigns. Although important to characterising a path, it does not directly provide any performance data.

Application can expose their delivery expectations to the network, by setting the Differentiated Services Code Point (DSCP) field of IPv4 and IPv6 packets. This can be used to inform network-layer queuing and forwarding, and can also provide information on the relative importance of packet information collected by measurement campaigns, but does not directly provide any performance data. This field provides explicit information that can be used in place of inferring traffic requirements (e.g., by inferring QoS requirements from port information via a multi-field classifier). The DSCP value can therefore impact the quality of experience for a flow. Observations of service performance need to consider this field when a network path has support for differentiated service treatment.

Explicit Congestion Notification (ECN) uses a codepoint in the network-layer header. Use of ECN can offer gains in terms of increased throughput, reduced delay, and other benefits when used over a path that includes equipment that supports an AQM method that performs Congestion Experienced (CE) marking of IP packets . This exposes the presence of congestion on a network path to the transport and network layer. The reception of Congestion Experienced (CE) marked packets can therefore be used to monitor the presence and estimate the level of incipient congestion on the upstream portion of the path from the point of observation (Section 2.5 of ). Because ECN marks carried in the IP protocol header, measuring ECN can be much easier than metering packet loss. However, interpreting the marking behaviour (i.e., assessing congestion and diagnosing faults) requires context from the transport layer (path RTT, visibility of loss - that could be due to queue overflow, congestion response, etc) . Some ECN-capable network devices can provide richer (more frequent and fine-grained) indication of their congestion state. Setting congestion marks proportional to the level of congestion (e.g., DCTP , and L4S ). AQM and ECN offer a range of algorithms and configuration options, it is therefore important for tools to be available to network operators and researchers to understand the implication of configuration choices and transport behaviour as use of ECN increases and new methods emerge . ECN-monitoring is expected to become important as AQM is deployed that supports ECN describes the transport layer feedback information that accompanies the use of ECN.

The common language between network operators and application/content providers/users is packet transfer performance at a layer that all can view and analyze. For most packets, this has been transport layer, until the emergence of QUIC, with the obvious exception of VPNs and IPsec. When encryption conceals more layers in a packet, people seeking understanding of the network operation need to rely more on pattern inferences and other heuristics. The accuracy of measurements therefore suffers, as does the ability to investigate and troubleshoot interactions between different anomalies. For example, the traffic patterns between server and browser are dependent on browser supplier and version, even when the sessions use the same server application (e.g., web e-mail access). Even when measurement datasets are made available (e.g., from endpoints) additional metadata, such as the state of the network, is often required to interpret the data. Collecting and coordinating such metadata is more difficult when the observation point is at a different location to the bottleneck/device under evaluation. Packet sampling techniques can be used to scale processing involved in observing packets on high rate links. This only exports the packet header information of (randomly) selected packets. The utility of these measurements depends on the type of bearer and number of mechanisms used by network devices. Simple routers are relatively easy to manage, a device with more complexity demands understanding of the choice of many system parameters. This level of complexity exists when several network methods are combined. This section discusses topics concerning transport measurement.

Often measurements can only be understood in the context of the other flows that share a bottleneck. A simple example is the monitoring of AQM. For example, FQ-CODEL , combines sub queues (statistically assigned per flow), management of the queue length (CODEL), flow-scheduling, and a starvation prevention mechanism. Usually such algorithms are designed to be self-tuning, but current methods typically employ heuristics that can result in more loss under certain path conditions (e.g., large RTT, effects of multiple bottlenecks ). In-network measurements that can distinguish between upstream and downstream metrics with respect to the measurement point. They are particularly useful for locating the source of problems or to assess the performance of a network segment or a particular device configuration.

Traffic measurements (e.g. Traffic volume, loss, latency) is used by operators to help plan deployment of new equipment and configurations in their networks. Data is also important to equipment vendors who need to understand traffic trends traffic and patterns of usage as inputs to decisions about planning products and provisioning for new deployments. This measurement information can also be correlated with billing information when this is also collected by an operator. A network operator supporting traffic that uses transport header encryption may not have access to per-flow measurement data. Trends in aggregate traffic can be observed and can be related this to the endpoint addresses being used, but it may not be possible to correlate patterns in measurements with changes in transport protocols (e.g., the impact of changes in introducing a new transport protocol mechanism). This increases the dependency on other indirect sources of information to inform planning and provisioning.

Traffic measurements (e.g., traffic volume, loss, latency) can be used by various actors to help understand the performance available to users of a network segment. While active measurements may be used in-network passive measurements can have advantages in terms of eliminating unproductive traffic, reducing the influence of test traffic on the overall traffic mix, and the ability to choose the point of measurement .

Transport header information is useful for a variety of operational tasks : to diagnose network problems, assess performance, capacity planning, management of denial of service threats, and responding to user performance questions. These tasks seldom involve the need to determine the contents of the transport payload, or other application details. A network operator supporting traffic that uses transport header encryption can see only encrypted transport headers. This prevents deployment of performance measurement tools that rely on transport protocol information. Choosing to encrypt all information may be expected to reduce the ability for networks to “help” (e.g., in response to tracing issues, making appropriate Quality of Service, QoS, decisions). For some this will be blessing, for others it may be a curse. For example, operational performance data about encrypted flows needs to be determined by traffic pattern analysis, rather than relying on traditional tools. This can impact the ability of the operator to respond to faults, it could require reliance on endpoint diagnostic tools or user involvement in diagnosing and troubleshooting unusual use cases or non-trivial problems. Although many network operators utilise transport information as a part of their operational practice, the network will not break because transport headers are encrypted.

Congestion control is a key transport function. Many network operators implicitly accept that TCP traffic to comply with a behaviour that is acceptable for use in the shared Internet. TCP algorithms have been continuously improved over decades, and they have reached a level of efficiency and correctness that custom application-layer mechanisms will struggle to easily duplicate . A standards-compliant TCP stack provides congestion control that is therefore judged safe for use across the Internet. Applications developed on top of well-designed transports can be expected to appropriately control their network usage, reacting when the network experiences congestion, by back-off and reduce the load placed on the network. This is the normal expected behaviour for TCP and other IETF-defined transports. Tools exist that can interpret the transport protocol header information to help understand the impact of specific transport protocols (or protocol mechanisms) on other traffic that shares their network. An observation in the network can gain understanding of the dynamics of a flow and its congestion control behaviour. Analysing observed packet sequence numbers can be used to help build confidence that an application flow backs-off its share of the network load in the face of persistent congestion, and hence to understand whether the behaviour is appropriate for sharing limited network capacity. For example, it is common to visualise plots of TCP sequence numbers versus time for a flow to understand how a flow shares available capacity, deduce its dynamics in response to congestion, etc.

UDP provides a minimal message-passing transport that has no inherent congestion control mechanisms. Because congestion control is critical to the stable operation of the Internet, applications and other protocols that choose to use UDP as an Internet transport must employ mechanisms to prevent congestion collapse, avoid unacceptable contributions to jitter/latency, and to establish an acceptable share of capacity with concurrent traffic . A network operator has no way of knowing the specific methods used by a UDP application, unless the header format can be determined. Tools are needed to understand if UDP flows comply with congestion control expectations and therefore whether there is a need to deploy methods such as rate-limiters, transport circuit breakers or other methods to enforce acceptable usage. UDP flows that expose a well-known header by specifying the format of header fields can allow information to be observed that gains understanding of the dynamics of a flow and its congestion control behaviour. For example, tools exist to monitor various aspects of the RTP and RTCP header information of real-time flows (see ).

By correlating observations at multiple points along the path (e.g., at the ingress and egress of a network segment), an observer can determine the contribution of a portion of the path to an observed metric (to locate a source of delay, jitter, loss, reordering, congestion marking, etc). Information provided by tools can help determine whether mechanisms are needed in the network to prevent flows from acquiring excessive network capacity. Operators can manage traffic flows (e.g., to prevent flows from acquiring excessive network capacity under severe congestion) by deploying rate-limiters, traffic shaping or network transport circuit breakers .

This section examines implications of encrypting specific transport header information.

Some transport information is made visible in the network-layer protocol header. These header fields are not encrypted and can be used to make flow observations. Endpoints should expose flow information in the IPv6 Flow Label in the network-layer header. This can be used to inform network-layer queuing, forwarding (e.g., for equal cost multi-path (ECMP) routing, and Link Aggregation (LAG)). For transport measurement, this can provide useful information to assign packets to flows in the data collected by measurement campaigns, but does not directly provide any performance data. Similarly the Differentiated Services CodePoint (DCSP) indicates expected forwarding treatment . The ECN field provides observable congestion data and can help inform measurement of flow congestion .

To measure and analyse a transport protocol, a measurement tool needs to be able to identify traffic flows. Aggregation of sessions, and persistent use of established transport flows by multiple sessions means that a flow at the transport layer is not necessarily the same as a flow seen at the application layer. This is usually not a consequence. Data is measured for the aggregate transport flow. Some measurement methods sample traffic, rather than collecting all packets passing through a measurement point. These methods still require a way to determine the presence, size and position of any observable header fields - but may need to do this without observing a protocol exchange for a connection setup.

If flow information is observed from transport headers, then there needs to be a way to identify the format of the header . Some IETF transport protocols are identified by an IP protocol number (e.g. ,TCP, SCTP, UDP). All IETF-defined transport protocols include a transport port field in their transport header. Higher layer protocols (e.g., HTTP) can be sometimes be observed by a well-known port value, which can be indicative of the protocol being encapsulated, but there is no way to enforce this usage. This can be used to configure decapsulation, alternatives include a "magic" number placed at the start of each UDP datagram. Once the protocol has been determined, the transport header can be determined from a published specification. If multiple formats are permitted, this may also require observing the protocol version being used and possibly parameter negotiation at connection setup.

When a transport is used as a substrate, the transport provides an encapsulation that allows another transport flow to be within the payload of a transport flow. The transported protocol header may provide additional information for multiplexing multiple flows over the same 5-tuple. The UDP Guidelines provides some guidance on using UDP as a substrate protocol. If there is no additional information about the protocol transported by the substrate, this may be viewed as an opaque traffic aggregate, and prevents transport measurement in the network. Examples include GRE-in-UDP , SCTP-in-UDP. The GRE-in-UDP encapsulation may encrypt the payload, but does not encrypt the GRE protocol header.

With the proliferation of mobile connected devices, there is a stated need for connection-oriented protocols to maintain connections after a network migration by an endpoint. The ability and desirability of in-network devices to track such migration depends on the context. On the one hand, a load-balancer device in front of server may find it useful to map a migrated connection to the same server endpoint. On the other hand, a user performing migration to avoid detection may prefer the network not to be able to correlate the different parts of a migrating session. Care must then be exercised to make sure that the information encoded by the endpoints is not sufficient to identify unique flows and facilitate a persistent surveillance attack vector . The impact of flow migration on measurement activities depends on the data being measured, rate of migration and level of encryption that is employed. Requirements for load balancing and mobility can lead to complex protocol interactions.

Transports can expose that start and end of flows in a transport header field (e.g., TCP SYN, FIN, RST). This can also help measurement devices identify the start of flows, or to remove stale flow information. This information is supplemental - flows can start and end at any time, the Internet network layer provides only a best effort service that allows alternate routing, reordering, loss, etc, so a network measurement tool can not rely upon observing these indicators. The time to complete a protocol connection and/or session setup can be reported. Flow information can provide in-network devices to manage their forwarding state . It can assist a firewall in deciding which flows are permitted through a security gateway, or to help maintain the network address translation (NAT) bindings in a NAPT or application layer gateway. This information may also find use in load balancers, where visibility of the 5-tuple could assist in selecting a server . Access to flow information and an observable start/stop indication can avoid stateful middleboxes relying on timeouts to remove old state. Without this, middleboxes are unaware when a particular flow ceases to be used by an application . This can lead to the state table entries keeping state for less time for flows that are not identifiable.

The TCP or RTP sequence number can be observed in one direction (the direction that carries data segments). An authenticated header prevents this field being modified or terminated/split by a network device, but allows this still to be used to observe progress of the network flow. An incrementing sequence number enables detection of loss (either by correlating ingress and egress value, or when assuming that all packets follow a single path), duplication and reordering (with understanding that not necessarily all packets of a flow follow the same path, and reordering can complicate processing of observations). Tools are widely available to interpret RTP and TCP sequence numbers, ranging from open source tools to dedicated commercial packages. As for TCP, use by in-network measurement devices needs to account for the impact of load-balancing of flows, changes in forwarding behaviour, measurement loss (rather than observed packet loss), etc.

Acknowledgement (ACK) data provides information about the path from the network device to the remote endpoint. The information can help identify packet loss (or the point of loss), RTT, and other network-related performance parameters (e.g., throughput, jitter, reordering). Unless this information is correlated with other data there is no way to disambiguate the cause of impairments (congestion loss, link transmission loss, equipment failure). An in-network device must not modify the flow of end-to-end ACK data when using an authenticated protocol. That is, must not use the in-network methods described in . This can impact the performance and/or efficiency (e.g., cost) of using paths where the return capacity is limited or has implications on the overall design (e.g., using TCP with cellular mobile uplinks, DOCSIS uplinks). The TCP stream can be observed by correlating the stream of TCP ACKs that flow from a receiver in the return direction. Although these ACKs are cumulative, and are not necessarily sent on the same path as the forward data, when visible, their sequence can confirm successful transmission and the path RTT. In the case of TCP they may also indicate packet loss (duplicate ACKs). An RTP session can provide reception information feedback using the RTCP framework. This reception information and can be observed by in-network measurement devices and can be interpreted to provide a variety of quality of experience information for the related RTP flow, as well as basic network performance data (RTT, loss, jitter, etc).

The use of timestamps for latency and jitter measurements is discussed in other sections of the current version of the document.

Transport protocols that use ECN need to provide ECN feedback information in the transport header to inform the sender whether packets have been received with an ECN CE-mark . This information can be in the form of feedback once each RTT or more frequent. The latter may involve sending a detailed list of all ECN-marked packets (e.g., and ). The detailed information can provide detail about the pattern and rate of marking. The information provided in these protocol headers can help a network operator to understand the congestion status of the forward path and the impact of marking algorithms on the traffic that is carried . IETF specifications for Congestion Exposure (CONVEX) is an example of a framework that monitors reception reports for CE-marked packets to support network operations.

This section is not complete - later revision may determine other fields or remove this section.

Understanding and analysing transport protocol behaviour typically demands tracking changes to the protocol state at the transport endpoints. Although protocols communicate state information in their protocol headers, a protocol implementation typically also contains internal state that is not directly visible from observing transport protocol headers. Effective measurement tools need to consider that not all packets may be observed (due to drops at the capture tap or because packets take an alternate route that does not pass the tap). Some flows of packets may also be encapsulated within a maintenance domain in other protocols, which further complicates analysis. Some examples of using network measurements of transport headers to infer internal TCP transport state information include: The TCP congestion window (cwnd) and slow start threshold (ssthresh). Tools for analysing in-network performance of TCP may observe sequence number to infer the current congestion controller state. The TCP RTT estimator and TCP Retransmission Time Out (RTO) value. This can be estimated by correlating sequence and acknowledgement numbers, or possibly by observing TCP timestamp options. Use of pacing (and pacing rate) and use of methods such as Proportional Rate Reduction (PRR) and Congestion Window validation (CWV). This may be estimated from observing timing of segments with TCP sequence numbers. This is important to some congestion control mechanisms and can be important for applications that are rate limited or send traffic bursts. Receiver window and flow control state. This may be inferred from information in TCP ACK segments. It is important to applications where the remote endpoint is resource constrained, or the path exhibits a large RTT. Retransmission state and receiver buffer. This may be inferred from information in TCP ACK segments (especially when SACK blocks are provided), this can be important to the performance of applications that send traffic bursts. Use of ACK delay and Nagle algorithm. This may be estimated from observing timing of segments with TCP sequence numbers, and is important to the performance of thin application flows.

Transport measurement introduces the following requirements to identify the observable information: Observable protocol type and version information is needed to identify the protocol being used when characterising the traffic, and to enable further observation of the flow. Observable format information is needed to allow an observer to determine the presence of any observable header fields. A published specification is needed to allow an observer to determine the size and position of any observable header fields so that these fields may be decoded by a measurement tool. Observable flow start/stop information can assist some forms of measurement and has utility for middleboxes that track state. The need for in-network transport measurement introduces the following requirements for observable information in transport header fields: Observable transport information to determine the progress of flows for each direction of communication. This requires observable packet numbers. Observable transport information to determine loss, and understand the response to congestion for a network segment. This requires observable reception information (e.g., packet acknowledgment information). Observable transport information is needed for more advanced measurement of latency, jitter, etc. This requires an observable field and a method to correlating return information with the observed field. This could utilise a packet number and/or transmission timestamp information. This information needs to be available in both directions of transmission. Exposure of Transport ECN feedback provides a powerful tool to understand ECN-enabled AQM-based networks. (Forward ECN information is already observable in the network header).

This section explores key implications of working with encrypted transport protocols.

Independent observation by multiple actors is important for scientific analysis. Encrypting transport header encryption changes the ability for other actors to collect and independently analyse data. Internet transport protocols employ a set of mechanisms. Some of these need to work in cooperation with the network layer - loss detection and recovery, congestion detection and congestion control, some of these need to work only end-to-end (e.g., parameter negotiation, flow-control). When encryption conceals information in the transport header, it could be possible for an applications to provide summary data on performance and usage of the network. This data could be made available to other actors. However, this data needs to contain sufficient detail to understand (and possibly reconstruct the network traffic pattern for further testing) and to be correlated with the configuration of the network paths being measured. Sharing information between actors needs also to consider the privacy of the user and the incentives for providing accurate and detailed information. Protocols that expose the state information used by the transport protocol in their header information (e.g., timestamps used to calculate RTT, packet numbers used to asses congestion and requests for retransmission) provide an incentive for the sending endpoint to provide correct information, increasing confidence that the observer understands the transport interaction with the network. This becomes important when considering changes to transport protocols, changes in network infrastructure, or the emergence of new traffic patterns.

If "unknown" or "uncharacterised" traffic patterns form a small part of the traffic aggregate passing through a network device or segment of the network the path, the dynamics of the uncharacterised traffic may not have a significant collateral impact on the performance of other traffic that shares this network segment. Once the proportion of this traffic increases, the need to monitor the traffic and determine if appropriate safety measures need to be put in place. Tracking the impact of new mechanisms and protocols requires traffic volume to be measured and new transport behaviours to be identified. This is especially true of protocols operating over a UDP substrate. The level and style of encryption needs to be considered in determining how this activity is performed. On a shorter timescale, information may also need to be collected to manage denial of service attacks against the infrastructure.

Attention therefore needs to be paid to the expected scale of deployment of new protocols and protocol mechanisms. Whatever the mechanism, experience has shown that it is often difficult to correctly implement combination of mechanisms . These mechanisms therefore typically evolve as a protocol matures, or in response to changes in network conditions, changes in network traffic or changes to application usage. The growth and diversity of applications and protocols using the Internet continues to expand - and there has been recent interest in a wide range of new transport methods, e.g., Larger Initial Window, Proportional Rate Reduction (PRR), congestion control methods based on measuring bottleneck bandwidth and round-trip propagation time, the introduction of AQM techniques and new forms of ECN response (e.g., Data Centre TCP, DCTP , and methods proposed for Low Latency Low Loss Scalable throughput, L4S). For each new method it is desirable to build a body of data reflecting its behaviour under a wide range of deployment scenarios, traffic load, and interactions with other deployed/candidate methods. Measurement therefore has a critical role in the design of transport protocol mechanisms and their acceptance by the wider community (e.g., as a method to judge the safety for Internet deployment. Open standards suggest that such evaluation needs to include independent observation and evaluation of performance data.

The transport layer provides the first end-to-end interactions across the Internet. Transport protocols are layered directly over the network service and are sent in the payload of network-layer packets. However, this simple architectural view hides one of the core functions of the transport – to discover and adapt to the properties of the Internet path that is currently being used. The design of Internet transport protocols is as much about trying to avoid the unwanted side effects of congestion on a flow and other capacity-sharing flows, avoiding congestion collapse, adapting to changes in the path characteristics, etc., as it is about end-to-end feature negotiation, flow control and optimising for performance of a specific application. To achieve stable Internet operations the IETF transport community, has to date, relied heavily on measurement and insight provided from the wider community to understand the trade-offs and to inform selection of select appropriate mechanisms to ensure a safe, reliable and robust Internet since the 1990's. There are many motivations for deploying encrypted transports, and encryption of transport payloads. The increasing public concerns about the interference with Internet traffic have led to a rapidly expanding deployment of encryption to protect end-user privacy, in protocols like QUIC. At the same time, network operators and access providers, especially in mobile networks, have come to rely on the in-network functionality provided by middleboxes both to enhance performance and support network operations. This document has expanded upon the expected implications on operational practices when working with encrypted transport protocols, and offers insight into the potential benefit of authentication, encryption and techniques that require in-network devices to interpret specific protocol header fields. It presents a need for architectural changes and consideration of approaches to the way network transport protocols are designed when using encryption. The use of encryption at the transport layer comes with implications that need to be considered: Encrypting all transport information eliminates the incentive for operators to troubleshoot what they cannot interpret: one flow experiencing packet loss looks like any other. When transport header encryption prevents decoding the transport header (if sequence numbers and flow ID are obscured), and hence understanding the impact on a particular flow or flows that share a common network segment. Encrypted traffic therefore implies "don't touch", and a likely first response will be "can't help, no trouble found", or the need to add complexity that comes with an additional operational cost . The use of transport header encryption may reduce the range of actors who can capture useful measurement data. This may in future restrict the information sources available to the Internet community to understand the operation of the network and transport protocols, necessary to inform standardisation and design decisions for new protocols, equipment and operational practices. There are dangers in a model where transport information is only observable at endpoints: i.e., at user devices and within service platforms and a need for independently captured data to develop open standards and stimulate research into new methods. Published transport specifications allow operators to check compliance. This can bring assurance to those operating networks, often avoiding the need to deploy complex techniques that routinely monitor and manage TCP/IP traffic flows (e.g. Avoiding the capital and operational costs of deploying flow rate-limiting and network circuit-breaker methods). This should continue when encrypted transport headers are used, but methods need to confirm that the traffic produced conforms to the expectations of the operator or developer. The use of encryption could make it harder to determine which transport methods are being used across a network segment and the trends in usage. This could impact the ability for an operator to anticipate the need for network upgrades and roll-out. It can also impact on-going traffic engineering activities. Although the impact in many case may be small, there are cases where operators directly support services (e.g., in radio links, or to troubleshoot QoS-related issues). The more complex the underlying infrastructure the more important this impact. An appropriate vantage point, coupled with timing information for the flow (fine-grained timestamps) is a valuable tool in benchmarking equipment/configurations and understanding non-trivial interactions. Encryption restricts the ability to explore interactions between functions at different protocol layers. This is a side-effect of not allowing a choice of the vantage point from which this information is observed. This can be important (e.g., in examining collateral impact of flows sharing a bottleneck, or where the intention is to understand the interaction between a layer 2 function (e.g., radio resource management policy, a channel impairment, an AQM configuration, a Per Hop Behaviour (PHB) or scheduling method, and a transport protocol). Since the introduction of congestion control, TCP has continued to be the predominate transport, with a consistent approach to avoiding congestion collapse. There is a risk that the diversity of transport mechanisms could also increase, with incentives to use a wide range of methods, this is not in itself a problem, nor is this a direct result of encryption. Encryption of all headers places the onus on validation in the hands of developers. While there is little to doubt that developers will seek to produce high quality code for their target use, it is not clear whether there is sufficient incentive to ensure good practice that benefits the wide diversity of requirements from the Internet community as a whole. The use of encryption needs to be weighed against the reduced visibility of the interactions between traffic, the network and the mechanisms. Especially, if a development cycle could focus on specific protocols/applications and then offer incentives for optimisations that could prove suboptimal for users or operators that utilise a network segments with different characteristics than targeted by the developer. The use of encryption may impede independent research into new mechanisms, measurement of behaviour, and development initiatives. Experience shows that transport protocols are complicated to design and complex to deploy, and that individual mechanisms need to be evaluated while considering other mechanism, across a broad range of network topologies and with attention to the impact on traffic sharing the capacity. Adopting pervasive encryption of transport information could eliminate the independent self-checks that have previously been in place from research and academic contributors (e.g., the role of the IRTF ICCRG, and research publications in reviewing new transport mechanisms and assessing the impact of their experimental deployment). Pervasive use of transport header encryption can impact the ways that future protocols are designed and deployed. The choice of whether candidate transport designs should encrypt their protocol headers therefore needs to be taken based not just on security considerations, but also on the impact on operating networks and the constrictions this may place on evolution of Internet protocols. While encryption of all transport information can help reduce ossification of the transport layer, it could result in ossification of the network service. There can be advantages in providing a level of ossification of the header in terms of providing a set of open specified header fields that are observable from in-network devices.

The author would like to thank all who have talked to him face-to-face or via email. ... This work has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 688421.The opinions expressed and arguments employed reflect only the authors' view. The European Commission is not responsible for any use that may be made of that information.

XX RFC ED - PLEASE REMOVE THIS SECTION XXX This memo includes no request to IANA.

This document is about design and deployment considerations for transport protocols. Authentication, confidentiality protection, and integrity protection are identified as Transport Features by RFC8095". As currently deployed in the Internet, these features are generally provided by a protocol or layer on top of the transport protocol; no current full-featured standards-track transport protocol provides these features on its own. Therefore, these features are not considered in this document, with the exception of native authentication capabilities of TCP and SCTP for which the security considerations in RFC4895. Like congestion control mechanisms, security mechanisms are difficult to design and implement correctly. It is hence recommended that applications employ well-known standard security mechanisms such as DTLS, TLS or IPsec, rather than inventing their own.

&RFC2119; &RFC3168; &RFC2474; &RFC3234; &RFC3135; &RFC8087; &RFC4585; &RFC5246; &RFC7525; &RFC4301; &RFC7567; &RFC7713; &RFC6437; &RFC7624; &RFC3449; &RFC4302; &RFC6347; &RFC8084; &RFC8085; &RFC8086; &RFC7258; &RFC4303; &RFC6679; &RFC7928; &RFC3819; &RFC5925; &RFC5559; &RFC3550; &RFC5236; &RFC4737; &I-D.mm-wg-effect-encrypt; &I-D.ietf-quic-transport; &I-D.dolson-plus-middlebox-benefits; &I-D.trammell-plus-abstract-mech; &I-D.trammell-plus-statefulness; &I-D.ietf-aqm-fq-codel; &I-D.ietf-aqm-codel; &I-D.ietf-aqm-pie; &I-D.ietf-tcpm-dctcp; &I-D.ietf-tcpm-accurate-ecn; &I-D.ietf-tsvwg-l4s-arch; &I-D.ietf-ippm-6man-pdm-option;

-00 This is an individual draft for the IETF community -01 This draft was a result of walking away from the text for a few days and then reorganising the content. -01 This draft fixes textua errors. Comments from the community are welcome on the text and recommendations.