]> China Mobile

China yaokehan@chinamobile.com

Huawei Technologies

China c.l@huawei.com

Telefonica

luismiguel.contrerasmurillo@telefonica.com

Qualcomm Europe, Inc.

jros@qti.qualcomm.com

Huawei Technologies

China shihang9@huawei.com

Routing Computing-Aware Traffic Steering CATS, metric Computing-Aware Traffic Steering (CATS) is a traffic engineering approach that optimizes the steering of traffic to a given service instance by considering the dynamic nature of computing and network resources. In order to consider the computing and network resources, a system needs to share information (metrics) that describes the state of the resources. Metrics from network domain have been in use in network systems for a long time. This document defines a set of metrics from the computing domain used for CATS. Discussion Venues Discussion of this document takes place on the Computing-Aware Traffic Steering Working Group mailing list (cats@ietf.org), which is archived at . Source for this draft and an issue tracker can be found at .

Introduction Service providers are deploying computing capabilities across the network for hosting applications such as distributed AI workloads, AR/VR and driverless vehicles, among others. In these deployments, multiple service instances are replicated across various sites to ensure sufficient capacity for maintaining the required Quality of Experience (QoE) expected by the application. To support the selection of these instances, a framework called Computing-Aware Traffic Steering (CATS) is introduced in . CATS is a traffic engineering approach that optimizes the steering of traffic to a given service instance by considering the dynamic nature of computing and network resources. To achieve this, CATS components require performance metrics for both communication and compute resources. Since these resources are deployed by multiple providers, standardized metrics are essential to ensure interoperability and enable precise traffic steering decisions, thereby optimizing resource utilization and enhancing overall system performance. Metrics from network domain have already been defined in previous documents, e.g., , ，and , and been in use in network systems for a long time. This document focuses on categorizing the relevant metrics at the computing domain for CATS into three levels based on their complexity and granularity.

Conventions and Definitions This document uses the following terms defined in :

• Computing-Aware Traffic Steering (CATS)
• Service
• Service contact instance

Design Principles

Three-Level Metrics As outlined in , the resource model that defines CATS metrics MUST be scalable, ensuring that its implementation remains within a reasonable and sustainable cost. Additionally, it MUST be useful in practice. To that end, a CATS system should select the most appropriate metric(s) for instance selection, recognizing that different metrics may influence outcomes in distinct ways depending on the specific use case. Introducing a definition of metrics requires balancing the following trade-off: if the metrics are too fine-grained, they become unscalable due to the excessive number of metrics that must be communicated through the metrics distribution protocol. (See for a discussion of metrics distribution protocols.) Conversely, if the metrics are too coarse-grained, they may not have sufficient information to enable proper operational decisions. Conceptually, it is necessary to define at least two fundamental levels of metrics: one comprising all raw metrics, and the other representing a simplified form---consisting of a single value that encapsulates the overall capability of a service instance. However, such a definition may, to some extent, constrain implementation flexibility across diverse CATS use cases. Implementers often seek balanced approaches that consider trade-offs among encoding complexity, accuracy, scalability, and extensibility. To ensure scalability while providing sufficient detail for effective decision-making, this document provides a definition of metrics that incorporates three levels of abstraction:

• Level 0 (L0): Raw metrics. These metrics are presented without abstraction, with each metric using its own unit and format as defined by the underlying resource.
• Level 1 (L1): Metrics normalized within categories. These metrics are derived by aggregating L0 metrics into multiple categories, such as network and computing. Each category is summarized with a single L1 metric by normalizing it into a value within a defined range of scores.
• Level 2 (L2): Single normalized metric. These metrics are derived by aggregating lower level metrics (L0 or L1) into a single L2 metric, which is then normalized into a value within a defined range of scores.

Level 0: Raw Metrics Level 0 metrics encompass detailed, raw metrics, including but not limited to:

• CPU: Base Frequency, boosted frequency, number of cores, core utilization, memory bandwidth, memory size, memory utilization, power consumption.
• GPU: Frequency, number of render units, memory bandwidth, memory size, memory utilization, core utilization, power consumption.
• NPU: Computing power, utilization, power consumption.
• Network: Bandwidth, capacity, throughput, bytes transmitted, bytes received, host bus utilization.
• Storage: Available space, read speed, write speed.
• Delay: Time taken to process a request.

L0 metrics serve as foundational data and do not require classification. They provide basic information to support higher-level metrics, as detailed in the following sections. L0 metrics can be encoded and exposed using an Application Programming Interface (API), such as a RESTful API, and can be solution-specific. Different resources can have their own metrics, each conveying unique information about their status. These metrics can generally have units, such as bits per second (bps) or floating point instructions per second (flops). Regarding network-related information, and define various performance metrics and their registries. Additionally, in , the ALTO WG introduced an extended set of metrics related to network performance, such as throughput and delay. For compute metrics, lists a set of cloud resource metrics.

Level 1: Normalized Metrics in Categories L1 metrics are organized into distinct categories, such as computing, communication, service, and composed metrics. Each L0 metric is classified into one of these categories. Within each category, a single L1 metric is computed using an aggregation function and normalized to a unitless score that represents the performance of the underlying resources according to that category. Potential categories include:

• Computing: A normalized value derived from computing-related L0 metrics, such as CPU, GPU, and NPU utilization.
• Communication: A normalized value derived from communication-related L0 metrics, such as communication throughput.
• Service: A normalized value derived from service-related L0 metrics, such as tokens per second and service availability
• Composed: A normalized value derived from an end-to-end aggregation function by levaraging both computing and communication metrics. For example, end-to-end delay computed as the sum of all delays along a path.

Editor note: detailed categories can be updated according to the CATS WG discussion. L0 metrics, such as those defined in , , , and , can be categorized into the aforementioned categories. Each category will employ its own aggregation function (e.g., weighted summary) to generate the normalized value. This approach allows the protocol to focus solely on the metric categories and their normalized values, thereby avoiding the need to process solution-specific detailed metrics.

Level 2: Single Normalized Metric. The L2 metric is a single score value derived from the lower level metrics (L0 or L1) using an aggregation function. Different implementations may employ different aggregation functions to characterize the overall performance of the underlying compute and communication resources. The definition of the L2 metric simplifies the complexity of collecting and distributing numerous lower-level metrics by consolidating them into a single, unified score. TODO: Some implementations may support the configuration of Ingress CATS-Forwarders with the metric normalizing method so that it can decode the information from the L1 or L0 metrics. Figure 1 provides a summary of the logical relationships between metrics across the three levels of abstraction.

Logic of CATS Metrics in levels

CATS Metrics Framework and Specification The CATS metrics framework is a key component of the CATS architecture. It defines how metrics are encoded and transmitted over the network. The representation should be flexible enough to accommodate various types of metrics along with their respective units and precision levels, yet simple enough to enable easy implementation and deployment across heterogeneous edge environments.

CATS Metric Fields This section defines the detailed structure used to represent CATS metrics. The design follows principles established in related IETF specifications, such as the network performance metrics outlined in . Each CATS metric is expressed as a structured set of fields, with each field describing a specific property of the metric. The following definition introduces the fields used in the CATS metric representations.

CATS Metric Fields

Next, we describe each field in more detail:

• Metric_Type (type): This field specifies the category or kind of CATS metric being measured, such as computational resources, storage capacity, or network bandwidth. It acts as a label that enables network devices to identify the purpose of the metric.
• Format (format): This field indicates the data encoding format of the metric, such as whether the value is represented as an integer, a floating-point number, or has no specific format.
• Format standard (format_std, optional): This optional field indicates the standard used to encode and decode the value field according to the format field. It is only required if the value field is encoded using a specific standard, and knowing this standard is necessary to decode the value field. Examples of format standards include ieee_754 and ascii. This field ensures that the value can be accurately interpreted by specifying the encoding method used.
• Length (length): This field indicates the size of the value field measured in octets (bytes). It specifies how many bytes are used to store the value of the metric. Examples include 4, 8, 16, 32, and 64. The length field is important for memory allocation and data handling, ensuring that the value is stored and retrieved correctly.
• Unit (unit): This field defines the measurement units for the metric, such as frequency, data size, or data transfer rate. It is usually associated with the metric to provide context for the value.
• Source (source, optional): This field describes the origin of the information used to obtain the metric. It may include one or more of the following non-mutually exclusive values:
  • 'nominal'. Similar to , "a 'nominal' metric indicates that the metric value is statically configured by the underlying devices. For example, bandwidth can indicate the maximum transmission rate of the involved device.
  • 'estimation'. The 'estimation' source indicates that the metric value is computed through an estimation process.
  • 'directly measured'. This source indicates that the metric can be obtained directly from the underlying device and it does not need to be estimated.
  • 'normalization'. The 'normalization' source indicates that the metric value was normalized. For instance, a metric could be normalized to take a value from 0 to 1, from 0 to 10, or to take a percentage value. This type of metrics do not have units.
  • 'aggregation'. This source indicates that the metric value was obtained by using an aggregation function.
  Nominal metrics have inherent physical meanings and specific units without any additional processing. Aggregated metrics may or may not have physical meanings, but they retain their significance relative to the directly measured metrics. Normalized metrics, on the other hand, might have physical meanings but lack units.
• Statistics (statistics, optional): This field provides additional details about the metrics, particularly if there is any pre-computation performed on the metrics before they are collected. It is useful for services that require specific statistics for service instance selection.
  • 'max'. The maximum value of the data collected over intervals.
  • 'min'. The minimum value of the data collected over intervals.
  • 'mean'. The average value of the data collected over intervals.
  • 'cur'. The current value of the data collected.
• Level (level): This field specifies the level at which the metric is measured. It is used to categorize the metric based on its granularity and scope. Examples include L0, L1, and L2. The level field helps in understanding the level of detail and specificity of the metric being measured.
• Value (value): This field represents the actual numerical value of the metric being measured. It provides the specific data point for the metric in question.

Aggregation and Normalization Functions In the context of CATS metric processing, aggregation and normalization are two fundamental operations that transform raw and derived metrics into forms suitable for decision-making and comparison across heterogeneous systems.

Aggregation Aggregation functions combine multiple metric values into a single representative value. This is particularly useful when metrics are collected from multiple sources or over time intervals. For example, CPU usage metrics from multiple service instances may be aggregated to produce a single load indicator for a service. Common aggregation functions include:

• Mean average: Computes the arithmetic average of a set of values.
• Minimum/maximum: Selects the lowest or highest value from a set.
• Weighted average: Applies weights to values based on relevance or priority.

The output of an aggregation function is typically a Level 2 metric, derived from multiple Level 0 metrics, or a level 2 metric, derived from multiple Level 0 or 1 metrics.

Aggregation function | | +------------+ | Aggregation | +----------+ ... | Function |---->| Metric 2 | +------------+ | | +----------+ | Metric 1.n |---->| | +------------+ +-------------------+ Input: Multiple values Output: Single value ]]>

Normalization Normalization functions convert metric values with or without units into unitless scores, enabling comparison across different types of metrics and systems. This is essential when combining metrics from a heterogeneous set of resources (e.g, latency measured in milliseconds with CPU usage measured in percentage) into a unified decision model. Normalization functions often map values into a bounded range, such as integers fron 0, to 5, or real numbers from 0 to 1, using techniques like:

• Sigmoid function: Smoothly maps input values to a bounded range.
• Min-max scaling: Rescales values based on known minimum and maximum bounds.
• Z-score normalization: Standardizes values based on statistical distribution.

Normalized metrics facilitate composite scoring and ranking, and can be used to produce Level 1 and Level 2 metrics.

Normalization function | Normalization Function |---->| Metric 2 | +----------+ +------------------------+ +----------+ Input: Value with or without units Output: Unitless value ]]>

On the Meaning of Scores in Heterogeneous Metrics Systems In a system like CATS, where metrics originate from heterogeneous resources---such as compute, communication, and storage---the interpretation of scores requires careful consideration. While normalization functions can convert raw metrics into unitless scores to enable comparison, these scores may not be directly comparable across different implementations. For example, a score of 4 on a scale from 1 to 5 may represent a high-quality resource in one implementation, but only an average one in another. This ambiguity arises because different implementations may apply distinct normalization strategies, scaling methods, or semantic interpretations. As a result, relying solely on unitless scores for decision-making can lead to inconsistent or suboptimal outcomes, especially when metrics are aggregated from multiple sources. To mitigate this, implementors of CATS metrics SHOULD provide clear and precise definitions of their metrics---particularly for unitless scores---and explain how these scores should be interpreted. This documentation should be designed to support operators in making informed decisions, even when comparing metrics from different implementations. Similarly, operators SHOULD exercise caution when making potentially impactful decisions based on unitless metrics whose definitions are unclear or underspecified. In such cases, especially when decisions are critical or sensitive, operators MAY choose to rely on Level 0 (L0) metrics with units, which typically offer a more direct and unambiguous understanding of resource conditions.

Level Metric Representations

Level 0 Metrics Several definitions have been developed within the compute and communication industries, as well as through various standardization efforts---such as those by the ---that can serve as L0 metrics. L0 metrics contain all raw metrics which are not considered to be standardized in this document, considering about their diversity and many other existing work. See Appendix A for examples of L0 metrics.

Level 1 Metrics L1 metrics are normalized from L0 metrics. Although they don't have units, they can still be classified into types such as compute, communication, service and composed metrics. This classification is useful because it makes L1 metrics semantically meaningful. The sources of L1 metrics is normalization. Based on L0 metrics, service providers design their own algorithms to normalize metrics. For example, assigning different cost values to each raw metric and do weighted summation. L1 metrics do not need further statistical values.

Normalized Compute Metrics The metric type of normalized compute metrics is "compute_norm", and its format is unsigned integer. It has no unit. It will occupy an octet. Example:

Example of a normalized L1 compute metric

Normalized Communication Metrics The metric type of normalized communication metrics is "communication_norm", and its format is unsigned integer. It has no unit. It will occupy an octet. Example:

Example of a normalized L1 communication metric

Normalized Composed Metrics The metric type of normalized composed metrics is "delay_norm", and its format is unsigned integer. It has no unit. It will occupy an octet. Example:

Example of a normalized L1 composed metric

Level 2 Metrics A Level 2 metric is a single-value, normalized metric that does not carry any inherent physical unit or meaning. While each provider may employ its own internal methods to compute this value, all providers must adhere to the representation guidelines defined in this section to ensure consistency and interoperability of the normalized output. Metric type is "norm_fi". The format of the value is unsigned integer. It has no unit. It will occupy an octet. Example:

Example of a normalized L2 metric

The single normalized value also facilitates aggregation across multiple service instances. When each instance provides its own normalized value, no additional statistical processing is required at the instance level. Instead, aggregation can be performed externally using standardized methods, enabling scalable and consistent interpretation of metrics across distributed environments.

Comparison among Metric Levels Metrics are progressively consolidated from L0 to L1 to L2, with each level offering a different degree of abstraction to address the diverse requirements of various services. Table 1 provides a comparative overview of these metric levels. Comparison among Metrics Levels

Level	Encoding Complexity	Extensibility	Stability	Accuracy
Level 0	High	Low	Low	High
Level 1	Medium	Medium	Medium	Medium
Level 2	Low	High	High	Medium

Since Level 0 metrics are raw and service-specific, different services may define their own sets---potentially resulting in hundreds or even thousands of unique metrics. This diversity introduces significant complexity in protocol encoding and standardization. Consequently, L0 metrics are generally confined to bespoke implementations tailored to specific service needs, rather than being standardized for broad protocol use. In contrast, Level 1 metrics organize raw data into standardized categories, each normalized into a single value. This structure makes them more suitable for protocol encoding and standardization. Level 2 metrics take simplification a step further by consolidating all relevant information into a single normalized value, making them the easiest to encode, transmit, and standardize. Therefore, from the perspective of encoding complexity, Level 1 and Level 2 metrics are recommended. When considering extensibility, Level 0 metrics allow new services to define their own custom metrics. However, this flexibility requires corresponding protocol extensions, and the proliferation of metric types can introduce significant overhead, ultimately reducing the protocol’s extensibility. In contrast, Level 1 metrics introduce only a limited set of standardized categories, making protocol extensions more manageable. Level 2 metrics go even further by consolidating all information into a single normalized value, placing the least burden on the protocol. Therefore, from an extensibility standpoint, Level 1 and Level 2 metrics are recommended. Regarding stability, Level 0 raw metrics may require frequent protocol extensions as new metrics are introduced, leading to an unstable and evolving protocol format. For this reason, standardizing L0 metrics within the protocol is not recommended. In contrast, Level 1 metrics involve only a limited set of predefined categories, and Level 2 metrics rely on a single consolidated value, both of which contribute to a more stable and maintainable protocol design. Therefore, from a stability standpoint, Level 1 and Level 2 metrics are preferred. In conclusion, for CATS, Level 2 metrics are recommended due to their simplicity and minimal protocol overhead. If more advanced scheduling capabilities are required, Level 1 metrics offer a balanced approach with manageable complexity. While Level 0 metrics are the most detailed and dynamic, their high overhead makes them unsuitable for direct transmission to network devices and thus not recommended for standard protocol integration.

Implementation Guidance on Using CATS Metrics <Authors Note: This part has been moved to , according to he chairs' sugguestion. Since this document is primarily on metric definition, rather than real implementations.>

Security Considerations TBD

IANA Considerations TBD

Informative References China Mobile Telefonica Huawei Technologies China Unicom Alibaba Group Distributed computing is a computing pattern that service providers can follow and use to achieve better service response time and optimized energy consumption. In such a distributed computing environment, compute intensive and delay sensitive services can be improved by utilizing computing resources hosted in various computing facilities. Ideally, compute services are balanced across servers and network resources to enable higher throughput and lower response time. To achieve this, the choice of server and network resources should consider metrics that are oriented towards compute capabilities and resources instead of simply dispatching the service requests in a static way or optimizing solely on connectivity metrics. The process of selecting servers or service instance locations, and of directing traffic to them on chosen network resources is called "Computing-Aware Traffic Steering" (CATS). This document provides the problem statement and the typical scenarios for CATS, which shows the necessity of considering more factors when steering traffic to the appropriate computing resource to better meet the customer's expectations. Huawei Technologies China Mobile Orange Telefonica Independent This document describes a framework for Computing-Aware Traffic Steering (CATS). Specifically, the document identifies a set of CATS components, describes their interactions, and provides illustrative workflows of the control and data planes. Nokia Bell Labs Telefonica Qualcomm Europe, Inc. Deutsche Telekom Service providers are starting to deploy computing capabilities across the network for hosting applications such as distributed AI workloads, AR/VR, vehicle networks, and IoT, among others. In this network-compute environment, knowing information about the availability and state of the underlying communication and compute resources is necessary to determine both the proper deployment location of the applications and the most suitable servers on which to run them. Further, this information is used by numerous use cases with different interpretations. This document proposes an initial approach towards a common exposure scheme for metrics reflecting compute and communication capabilities. n.d. n.d. This document defines the format for the IANA Registry of Performance Metrics. This document also gives a set of guidelines for Registered Performance Metric requesters and reviewers. This memo defines the set of initial entries for the IANA Registry of Performance Metrics. The set includes UDP Round-Trip Latency and Loss, Packet Delay Variation, DNS Response Latency and Loss, UDP Poisson One-Way Delay and Loss, UDP Periodic One-Way Delay and Loss, ICMP Round-Trip Latency and Loss, and TCP Round-Trip Delay and Loss. The cost metric is a basic concept in Application-Layer Traffic Optimization (ALTO), and different applications may use different types of cost metrics. Since the ALTO base protocol (RFC 7285) defines only a single cost metric (namely, the generic "routingcost" metric), if an application wants to issue a cost map or an endpoint cost request in order to identify a resource provider that offers better performance metrics (e.g., lower delay or loss rate), the base protocol does not define the cost metric to be used. This document addresses this issue by extending the specification to provide a variety of network performance metrics, including network delay, delay variation (a.k.a. jitter), packet loss rate, hop count, and bandwidth. There are multiple sources (e.g., estimations based on measurements or a Service Level Agreement) available for deriving a performance metric. This document introduces an additional "cost-context" field to the ALTO "cost-type" field to convey the source of a performance metric.

Appendix A

Level 0 Metric Representation Examples Several definitions have been developed within the compute and communication industries, as well as through various standardization efforts---such as those by the ---that can serve as L0 metrics. This section provides illustrative examples.

Compute Raw Metrics This section uses CPU frequency as an example to illustrate the representation of raw compute metrics. The metric type is labeled as compute_CPU_frequency, with the unit specified in GHz. The format should support both unsigned integers and floating-point values. The corresponding metric fields are defined as follows:

An Example for Compute Raw Metrics

Communication Raw Metrics This section takes the total transmitted bytes (TxBytes) as an example to show the representation of communication raw metrics. TxBytes are named as "communication type_TxBytes". The unit is Mega Bytes (MB). Format is unsigned integer or floating point. It will occupy 4 octets. The source of the metric is "Directly measured" and the statistics is "mean". Example:

An Example for Communication Raw Metrics

Delay Raw Metrics Delay is a kind of synthesized metric which is influenced by computing, storage access, and network transmission. Usually delay refers to the overal processing duration between the arrival time of a specific service request and the departure time of the corresponding service response. It is named as "delay_raw". The format should support both unsigned integer or floating point. Its unit is microseconds, and it occupies 4 octets. For example:

An Example for Delay Raw Metrics

Contributors Orange

mohamed.boucadair@orange.com

China Mobile

duzongpeng@chinamobile.com