| TAPS Working Group | P. Tiesel |
| Internet-Draft | T. Enghardt |
| Intended status: Informational | TU Berlin |
| Expires: May 7, 2019 | November 03, 2018 |
Communication Units Granularity Considerations for Multi-Path Aware Transport Selection
draft-tiesel-taps-communitgrany-03
This document provides an approach how to reason about the composition of multi-path aware transport stacks. It discusses how to compose the functionality needed by stacking existing internet protocols and the fundamental mechanisms that are used in multi-path systems and the consequences of applying them to different granularities of communication units, e.g, on a message or stream granularity. This document is targeted as guidance for automation of destination selection, path selection, and transport protocol selection.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on May 7, 2019.
Copyright (c) 2018 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
Today’s Internet architecture faces a communication endpoint with a set of choices, including choosing a transport protocol and picking an IP protocol version. In many cases, e.g., when fetching data from a CDN, an endpoint has also the choice of which endpoint instance, [I-D.brunstrom-taps-impl] calls these instances Derived Endpoint, to contact as DNS can return multiple alternative addresses.
If endpoints want to take advantage of multiple available paths, there is another bunch of, partially interdependent, choices:
Implementing an heuristic or strategy for choosing from this overwhelming set of transport options by each application puts a huge burden on the application developer. Thus, the decisions regarding all transport options mentioned so far should be supported and, if requested by the application, automated within a the transport layer. In order to build such automatization, we need to be able to compare the product of all transport options (destinations, paths, transport protocols and protocol options) available to choose the most appropriate.
As the protocols to be used are not known a priori and can differ depending on other transport options, this reasoning has to be independent of a specific protocol or implementation and allow to compare them even if they operate on different communication unit granularities.
When reasoning about network systems, layering traditionally has been the main guidance on where functionality is placed. Looking at modern systems, the classical concept of layers and their mapping to protocols becomes blurry. Protocols can operate on different granularities of communication units, i.e., the semantic units such as messages that the protocols distinguish. These communication units often do not match the PDUs used by the protocols, e.g., TCP segments do not necessarily align with messages at the application layer.
In this document, we do not want to take a protocol-centric perspective, but we focus on mechanisms a multi-access system is composed of and the communication units they operate on. This has several advantages:
Overall, this perspective allows us to compare mechanism like distributing requests of an application among different paths, MPTCP and using bandwidth aggregation proxies (as discussed within the IETF in the BANANA working group) despite their different nature and layer of implementation.
These communication units definitions are primarily used for reasoning about automatic stack composition. Therefore, depending on the protocol stack instance, a communication unit can span multiple protocol instances.
Some of these hierarchy levels correspond to objects in [I-D.ietf-taps-minset], but in case of Association and Association Set, we have to split categories as they may indeed be separate on the transport. Note the naming confusion concerning the term “flow” deriving from different perspective.
We also annotate the corresponding terminology used in [I-D.ietf-taps-arch] if it differs from the one used in this document.
An Message is a piece of data that has a meaning for the application. It is the smallest communication unit that we consider.
[I-D.ietf-taps-minset] correspondent: Message
Examples:
A Stream is an ordered sequence of related Messages that should be treated the same by the transport system.
[I-D.ietf-taps-minset] correspondent: Flow
Examples:
An Association multiplexes a set of Messages or Streams within the same Flow with common source and destination. Therefore these communication units become indistinguishable for the network. Association and flow describe the same concept, the former from the perspective of the application, the latter from the perspective of the network.
[I-D.ietf-taps-minset] correspondent: Flow-Group
[I-D.ietf-taps-arch] correspondent: Connection Group
Examples:
An Association Set or Flow Set is a set of Associations or Flows that belong together from an application point of view.
[I-D.ietf-taps-minset] correspondent: Flow-Group
Examples:
Transport protocols on the Internet provide a large variety of functionality. While the functionality of simple protocols like UDP is easy to describe (multiplexing streams of messages), describing the functionality of complex protocols such as QUIC, MPTCP or SCTP is manyfold as these protocols provide a set of commonly used functionality. Also, the same functionality can be provided at many places throughout the whole stack. In the following, we explore the set of functionality that can be provided by transport protocols.
Destination Selection refers to selecting one of multiple different destinations. This mechanism is applicable to any kind of communication unit and can occur on all layers.
Typical cases for destination selection include:
In practice, destination address selection is often tied to name resolution. As name resolution relies on both local decisions on the endpoint as well as decisions within the DNS infrastructure, this mechanism spreads across different administrative domains which each independently contribute to the overall selection result.
Path Selection refers to choosing which of the available paths to use. and can occur on the network layer and any layer below.
While path selection involves a choice of access network it does not need knowledge of or changes to the routing choices within the core network.
When doing path selection on small communication units like TCP segments, it is not uncommon to split path selection into two subproblems: Candidate Path Selection determines feasible and preferred choices, e.g., in case of MPTCP by establishing subflows. Afterwards, Per-Chunk Path Selection selects among these alternatives for each chunk. Thus, the first can be more expensive while the latter should be easy to execute.
TODO: Discuss difference between Multiple Provisioning Domains [RFC7556] or multiple access networks within the same provisioning domain – especially when it comes to integrating 3GPP mechanisms like [RFC5555] or [RFC7864].
Chunking refers to splitting an message, a stream or a set of associations into one or more parts. Typically, chunking splits only large messages or streams into multiple ones while keeping smaller entities untouched. Associations or Flows are typically not split, but sets of Associations or Flows might be partitioned. Once split into chunks, each chunk can be transferred individually over different transfer options.
Chunking can and does occur at different layers within a system:
In practice, chunking is often constrained in order to maintain certain properties that are desirable for the overall system. Examples such restrictions include the following:
Scheduling refers to distributing chunks or sets of chunks across multiple pre-chosen path. Thus, depending on the objectives, it can make sense to see scheduling as is nothing else than per-chunk path selection as defined above. In other cases, e.g. when trying to balance traffic, it makes sense to look at scheduling as a concept itself that uses chunking and per-chunk path selection as sub-mechanisms.
Examples of scheduling strategies include:
Transport option selection mechanisms are often intertwined. Which mechanism is used by which layer or which network component depends on the transfer objectives as well as the state of the network, e.g., availability, path throughput, path RTT, server load.
The cost and complexity of transport option selection depends on the network state used and the number of transfer options. If the transfer option selection only uses local state e.g., link availability, and the mechanism is predetermined and/or uses simple mechanisms, e.g., a simple hash function, the cost can even be negligible. An example where transfer option selection is cheap is ECMP within a router. In other cases, the cost can be non-trivial, e.g. when the selection involves queries to remote entities or even active network performance measurements. Such examples include DNS or DHT lookups, as used by some file sharing protocols, or network measurements like RTT and bandwidth estimations used by many video streaming applications. Indeed, costs may be prohibitive, e.g when requiring multiple DNS lookups for every 1 second chunk of a 20 minute video.
It may become necessary to take path layer components (middle-boxes) into account that interfere with the transport layer.
While the classical “End-To-End Arguments in System Design” [End-To-End] advocates for a dumb network and placing functionality as close to the edge and up in the stack as possible, there are always tussles of moving functionality up or down the stack. This document does not argue against pushing some multi-path functionality down the stack, but advocates to maintain the control of the overall system composition at the end host. Functionality provided by a path can indeed be a reason to choose this path for a given communication unit.
Some flow off-loading mechanisms that come in gestalt of of logical interfaces, e.g., [RFC7847]. These interfaces treat some association sets differently, which can be considered on-path functionality.
| Protocol | Congestion Control | Ordering | Reliability | Integrity P. | Confidentiality P. | Authenticity P. | Chunking | Multiplexing |
|---|---|---|---|---|---|---|---|---|
| HTTP | r | r | r | bytes | requests | |||
| HTTPS | r | r | r | r | r | r | bytes | requests |
| XMPP | r | r | r | (r) | (r) | (r) | messages | |
| SIP | + | (r) | (r) | (r) | messages | |||
| DTLS | + | + | + | services,name | ||||
| TLS | r | r | + | + | + | services,name | ||
| RTP | +(prf) | +(prf) | messages(prf) | messages | ||||
| SRTP | +(prf) | +(prf) | + | + | r(sig) | messages(prf) | messages | |
| QUIC | + | + | + | + | + | +(tls) | bytes | connection-id,+(tls) |
| UDP | ports | |||||||
| DCCP | + | ports | ||||||
| TCP | + | + | + | bytes | ports | |||
| MPTCP | + | + | + | bytes | ports | |||
| SCTP | + | + | + | bytes | ports,streams | |||
| IPsec (ESP) | + | + | r(ike) | spi,next-header | ||||
| IPsec (AH) | + | r(ike) | spi,next-header | |||||
| IP | (+(fr)) | (fragments) | address,next-header | |||||
| NEMO/IFOM | r | r | assoc. |
Legend:
fr: :Only when fragmentation is used and only to re-assemble IP PUDs
This work has been supported by Leibniz Prize project funds of DFG - German Research Foundation: Gottfried Wilhelm Leibniz-Preis 2011 (FKZ FE 570/4-1).