]> School of Cyber Science and Engineering, Sichuan University

No.24 South Section 1, Yihuan Road Chengdu 610000 China kaigao@scu.edu.cn

Application ALTO This document introduces an extension to the Application-Layer Traffic Optimization (ALTO) protocol, which enables announcements of the composition modes of multiple cost map services. Specifically, the composition mode defines how the results of multiple cost map services are combined to get the final prediction between two network endpoints. This extension allows ALTO servers to improve the accuracy of the prediction model at similar map sizes, and to efficiently enable differentiated services.

Introduction The Application-Layer Traffic Optimization (ALTO) protocol provides abstractions for application operators and/or end users to query network distance or property information. Specifically, ALTO has defined network map and cost map, which typically are used together, to provide a prediction model of distance information between endpoints in a network. Given the scale of the Internet today, it is unlikely that the prediction model can overfit. Thus, with higher model complexity, an ALTO service tends to provide better accuracy from the same implementation method. As a consequence, operators of the ALTO maps have to make the trade-off between service quality (accuracy of the predicated value) and model complexity (sizes of the maps). Currently, there is no standard way of composing the prediction results from multiple ALTO cost maps. Clients either only request a single pair of network and cost maps, or blindly select ALTO maps and compose the results. These approaches either make inefficient trade-offs, i.e., achieving substantial lower accuracy gains than occupied map sizes, or make incorrect use of the servers' exposed maps, i.e., the composition mode is different from how the server internally constructs the models. This extension is motivated by the ensemble method in machine learning . Ensemble method uses multiple prediction models to improve the "efficiency" and can typically achieve higher accuracy with the same model complexity. When the models are composed (or "ensembled") using the boosting method , models are ordered and higher-order models are trained not directly with the samples but residuals (prediction errors) of lower-order models. Thus, model accuracy and model complexity typically grow simultaneously with the number of models -- in the context of ALTO, the number of maps. Thus, an ALTO server may realize differentiated service by controlling the access to higher-order maps. Specifically, this extension defines a new type of ALTO resource called ALTO composition advertisement . The resource specifies the list of ALTO cost maps and how they are intended to be composed.

Conventions and Terminology 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 [RFC2119]. All numeric values are in network byte order. Values are unsigned unless otherwise indicated. Literal values are provided in decimal or hexadecimal as appropriate. Hexadecimal literals are prefixed with "0x" to distinguish them from decimal literals. This document reuses the terms defined in RFC 7285 .

Composition Modes This document has some requirements on the cost maps that can be composed. For cost maps that satisfy these requirements, 3 different composition modes are specified to define how the results of these maps must be combined.

Basic Requirements This extension has the following requirements: First, the cost maps to be composed must support a common cost type. Second, the prediction using a network map and a cost map must follow the same process. Specifically, for a given pair of source and destination network hosts (identified by their IP addresses), the prediction result must be computed as follows:

1. Find the source PID with the longest matching prefix for the source host.
2. Find the destination PID with the longest matching prefix for the destination host.
3. The prediction result is the distance between the source PID and the destination PID.

Composition Mode and Result Ensembling

All This composition mode is indicated by the string "all". If the composition mode is "all", for each source and destination hosts, the client MUST compute the (weighted) sum of the prediction results from each cost map and its corresponding network map. This mode implies that missing the prediction result of any cost map may lead to substantial prediction error.

Random This composition mode is indicated by the string "random". If the composition mode is "random", the client MAY obtain a prediction result by computing the (weighted) average of prediction results from any non-empty subset of the cost maps. This mode typically implies that the maps are generated using a bagging method, e.g., random forests.

Gradient This composition mode is indicated by the string "gradient". If the composition mode is "gradient", the client MUST interpret the cost maps as an ordered list and MAY obtain a prediction result by computing the (weighted) sum of the first K maps, where K is an arbitrary number that is no less than 1 and no greater than the number of cost maps. This mode typically implies that the maps are generated using a boosting method. It must be noted that prediction results of higher-order maps are useless without the results of lower-order maps in this mode.

ALTO Composition Advertisement

Media Type The composition advertisement resource is a virtual resource and the media type is only used to identify the type of the resource. The "media-type" field in its IRD entry MUST be "application/alto-composition+json".

HTTP Method The composition advertisement resource is a virtual resource and does not accept any HTTP method.

Accept Input Parameters None.

Capabilities The capabilities of a composition advertisement is a JSON object of type CompAdvCapabilities: ; [JSONNumber weights<1..*>;] } CompAdvCapabilities; ]]> with fields: comp-mode: ~ A JSONString whose value MUST either be "all", "random" or "gradient", as introduce in . cost-type-names: ~ A list of cost type names. Each cost type name MUST appear in the "cost-types" field in the "meta" field of the IRD, and MUST appear in the "cost-type-names" of each cost map whose resource ID is in the entry's "uses" field of the composition advertisement resource. The cost mode of this cost type MUST be "numerical". weights: ~ An optional list of weight coefficient for each cost map in the "uses" field of this resource. The length of this option MUST be equal to the length of the "uses" field.

Uses The resource ID of each cost map that may be composed as instructed by the capabilities of this resource.

Response None.

References Normative References Applications using the Internet already have access to some topology information of Internet Service Provider (ISP) networks. For example, views to Internet routing tables at Looking Glass servers are available and can be practically downloaded to many network application clients. What is missing is knowledge of the underlying network topologies from the point of view of ISPs. In other words, what an ISP prefers in terms of traffic optimization -- and a way to distribute it. The Application-Layer Traffic Optimization (ALTO) services defined in this document provide network information (e.g., basic network location structure and preferences of network paths) with the goal of modifying network resource consumption patterns while maintaining or improving application performance. The basic information of ALTO is based on abstract maps of a network. These maps provide a simplified view, yet enough information about a network for applications to effectively utilize them. Additional services are built on top of the maps. This document describes a protocol implementing the ALTO services. Although the ALTO services would primarily be provided by ISPs, other entities, such as content service providers, could also provide ALTO services. Applications that could use the ALTO services are those that have a choice to which end points to connect. Examples of such applications are peer-to-peer (P2P) and content delivery networks. Informative References Oregon State University Stanford University