| Internet-Draft | CATS Metric | October 2024 |
| Kehan, et al. | Expires 24 April 2025 | [Page] |
This document defines the computing metrics used in Computing-Aware Traffic Steering.¶
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 24 April 2025.¶
Copyright (c) 2024 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 Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
Many modern computing services are deployed in a distributed way. In this deployment mode, multiple service instances are deployed in multiple sites to provide equivalent function to end users. In order to provide better service to end users, a framework called CATS (Computing-Aware Traffic Steering) [I-D.ietf-cats-framework] is proposed.¶
CATS (Computing-Aware Traffic Steering) [I-D.ietf-cats-framework] is a traffic engineering approach that takes into account the dynamic nature of computing resources and network state to optimize service-specific traffic forwarding towards a given service contact instance. Various relevant metrics may be used to enforce such computing-aware traffic steering policies.¶
To effectively steer traffic to the appropriate service instance, network devices need a model of the service instance's computing status. A common definition of computing metrics is essential for effective coordination between network devices and computing systems. Without standardized computing metrics, devices on the network may interpret and respond to traffic conditions and computing load differently, leading to inefficiencies and potential conflicts. A standardized metric allows both network devices and computing systems to evaluate load consistently, enabling precise traffic steering decisions that optimize resource utilization and improve overall system performance.¶
Various considerations for metric definition are proposed in [I-D.du-cats-computing-modeling-description], which are useful in defining computing metrics.¶
Based on the considerations defined in [I-D.du-cats-computing-modeling-description], this document defines relevant computing metrics for CATS by categorizing the metrics into three levels based on their complexity and richness.¶
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.¶
This document uses terms defined in [I-D.ietf-cats-framework]. We list them below for clarification.¶
Computing-Aware Traffic Steering (CATS): An architecture that takes into account the dynamic nature of computing resources and network state to steer service traffic to a service instance. This dynamicity is expressed by means of relevant metrics.¶
Service: An offering that is made available by a provider by orchestrating a set of resources (networking, compute, storage, etc.).¶
Service instance: An instance of running resources according to a given service logic.¶
Many metrics are being discussed and/or defined in routing and computing area. Definition and usage of specific metrics are highly related to the use case, especially in IT use cases. However, when considering distributing compute metrics to network devices, appropriate categorizing and abstraction is required in order to not introduce extra complexity into the network.¶
Based on the abstraction level of metrics, this document defines three levels of metric to meet different requirements of different use cases:¶
Level 0(L0): Raw Metrics. In this level, the metrics are not abstracted, so different metrics use their own unit and format.¶
Level 1(L1): Normalized Metrics in Categories. In this level, the metrics are categorized into multiple dimensions, such as network, computing and storage. Each category metric is normalized into a value.¶
Level 2(L2): Fully Normalized Metric. In this level, metrics are normalized into a single value, the category information or raw metrics information cannot be interpreted from the value directly.¶
The metrics without any abstraction are Level 0 metrics. Therefore, Level 0 metrics encompass detailed, raw metrics, including but not limit to:¶
CPU: Base Frequency, Number of Cores, Boosted Frequency, Memory Bandwidth, Memory Size, Memory Utilization Ratio, Core Utilization Ratio, Power Consumption.¶
GPU: Frequency, Number of Render Unit, Memory Bandwidth, Memory Size, Memory Utilization Ratio, Core Utilization Ratio, Power Consumption.¶
NPU: Computing Power, Utlization Ratio, Power Consumption.¶
Network: Bandwidth, TXBytes, RXBytes, HostBusUtilization.¶
Storage: Available Space, Read Speed, Write Speed.¶
Delay: Time takes to process a request.¶
In L0, detailed information of a metric can be encoded into the protocol, and different service has its own metrics with different information elements. This kind of metrics are used widely in IT systems.¶
Regarding network related raw metrics, IPPM WG has defined many types of metrics in [performance-metrics]. [RFC9439] also defines a lot of metrics of packet performance and Throughput/Bandwidth. Regarding computing metrics, [I-D.rcr-opsawg-operational-compute-metrics] defines a set of cloud resource metrics.¶
In Level 1, the metrics will be categorized into different categories, and appropriate abstraction will be applied to each category. The Level 0 raw metrics can be categorized into multiple categories, such as computing, networking, storage and delay. In each category, the metrics are normalized into a value that present the state of the resource, making it as a Level 1 metric. Potential categories are shown below:¶
Computing: A normalized value generating from the computing related L0 metrics, such as CPU/GPU/NPU L0 metrics¶
Networking: A normalized value generating from the network related L0 metrics.¶
Storage: A normalized value generating from the storage L0 metrics.¶
Delay: A normalized value generating from computing/networking/storage metrics, reflecting the processing delay of a request.¶
Editor note: detailed categories can be updated according to the CATS WG discussion.¶
The L0 metrics, such as the ones defined in [performance-metrics] ,[RFC9439] and [I-D.rcr-opsawg-operational-compute-metrics] can be categorized into above categories. Each category will use its own method(weighted summary, etc.) to generate the normalized value. In this way, the protocol only care about the metric categories and its normalized value, and avoid to process the detailed metrics.¶
L2 metric is a one-dimensional value derived from a weighted sum of L1 metrics or from L0 metrics directly. Different service has its own normalization method which might use different metrics with different weight. For the ingress CATS router, it can compare the metric value to make the traffic steering decision (e.g., larger value has higher priority) . In some cases, some implementations may support to configure the ingress CATS router to know the metric normalizing method so that it can decode the affection from the L1 or L0 metrics.¶
This method simplifies the complexity of transmission and management of multiple metrics by consolidating them into a single, unified measure.¶
The below figure 1 shows the logic of metrics in Level 0, level 1 and level 2.¶
+--------------+
Level 2 +------| Normalized M |-------+
| +--------------+ |
| | |
| | Normalizing |
+---------+ +--------+ +--------+
Level 1 | Cate M1 | | Cate M2| | Cate M3| ...
+---------+ +--------+ +--------+
| | | | |
| | |Normalizing | |
+------+ +------+ +------+ +------+ +------+
Level 0 |Raw M1| |Raw M2|...|Raw M3|...|Raw M4| |Raw M5| ...
+------+ +------+ +------+ +------+ +------+
A hierarchical view of metrics has been shown in the section above. In this section, the detailed representation of metrics will be described.¶
[RFC9439] gives a good way to show the representation of some network metrics which is used for network capabilities exposure to applications. This document further describe the representation of CATS metrics.¶
Basically, in each metric level and for each metric, there will be some common fields for representation, including metric type, unit, and precision. Metric type is a name for network devices and protocols to recognize what the metric is. unit and precision are necessary for metric descripition. How many bits a metric occupies in protocols is also required.¶
Beyond these basic representations, the source of the metrics MUST also be declared. As defined in [RFC9439], there are three cost-sources, nominal, sla, and estimation. This document further divide the estimation type into three sub-types, direct measurement, aggregation, and normalization, since different levels of metrics require different sources to acquire CATS metrics. Directly measured metrics have physical meanings and units without any processing. Aggregation metrics can be either physically meaningful or not, and they maintain their meanings compared to the directly measured metrics. Normalized metrics can have physical meanings or not, but they do not have units, and they are just numbers that used for routing decision making.¶
To be more fine grained, This document refer to the definition of [RFC9439] on the metrics statistics.¶
Raw metrics have exact physical meanings and units. They are directly measured from the underlying computing resources providers. Lots of definition on this level of metrics have been defined in IT industry and other standardisations[DMTF], and this document only show some examples for different categories of metrics for reference.¶
The metric type of storage resources like SSD are named as “storage_type: SSD”. The storage space unit is megaBytes(MBs). Format is integer. It will occupy 2 octets. The unit of read or write speed is denoted as MB per second.¶
Example[TBA].¶
The metric type of network resources like bandwidth are named as “network_type: Bandwidth”. The unit is gigabits per second(Gb/s). Format is integer. It will occupy 2 octets. The unit of TXBytes and RXBytes is denoted as MB per second.¶
Example[TBA].¶
Delay is a kind of synthesized metric which is influenced by computing, storage access, and network transmission. It is named as “delay_raw”. Format should support integer and FP8. Its unit is microsecond. It will occupy 4 octets.¶
The sources of L0 metrics can be nominal, directly measured, or aggregated. Nominal L0 metrics are provided initially by resource providers. Dynamic L0 metrics are measured and updated during service stage. L0 metrics also support aggregation, in case that there are multiple service instances.¶
The statistics of L0 metrics will follow the definition of section 3.2 of [RFC9439].¶
Normalized metrics in categories have physical meanings but they do not have unit. They are numbers after some ways of abstraction, but they can represent their type, in case that in some use cases, some specific types of metrics require more attention.¶
The metric type of normalized compute metrics is “compute_norm”, and its format is integer. It has no unit. It will occupy a octet.¶
Example[TBA].¶
The metric type of normalized compute metrics is “storage_norm”, and its format is integer. It has no unit. It will occupy a octet.¶
Example[TBA].¶
The metric type of normalized compute metrics is “network_norm”, and its format is integer. It has no unit. It will occupy a octet.¶
Example[TBA].¶
The metric type of normalized compute metrics is “delay_norm”, and its format is integer. It has no unit. It will occupy a octet.¶
Example[TBA].¶
The sources of L1 metrics is normalized and support aggregation. 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 summation. L1 metric do not need further statistical values.¶
The fully normalized metric is a single value which does not have any physical meaning or unit. Each provider may have its own methods to derive the value, but all providers MUST follow the definition in this section to represent the fully normalized value.¶
Metric type is “Norm_fi”. The format of the value is non-negative integer. It has no unit. It will occupy a octet.¶
The fully normalized value also supports aggregation when there are multiple service instances providing these fully normalized values. When providing fully normalized values, service instances do not need to do further statistics.¶
From L0 to L1 to L2, the computing metric is consolidated. Different level of abstraction can meet the requirements from different services. Table 1 shows the comparison among metric levels.¶
| Level | Encoding Complexity | Extensibility | Stability | Accuracy |
|---|---|---|---|---|
| Level 0 | Complicated | Bad | Bad | Good |
| Level 1 | Medium | Medium | Medium | Medium |
| Level 2 | Simple | Good | Good | Medium |
Since Level 0 metrics are raw metrics, therefore, different services may have their own metrics, resulting in hundreds or thousands of metrics in total, this brings huge complexity in protocol encoding and standardization. Therefore, this kind of metrics are always used in customized IT systems case by case. In Level 1 metrics, metrics are categorized into several categories and each category is normalized into a value, therefore they can be encoded into the protocol and standardized. Regarding the Level 2 metrics, all the metrics are normalized into one single metric, it is easier to be encoded in protocol and standardized. Therefore, from the encoding complexity aspect, Level 2 and Level 1 metrics are suggested.¶
Similarly, when considering extensibility, new services can define their own new L0 metrics, which requires protocol to be extended as needed. Too many metrics type can create a lot of overhead to the protocol resulting in a bad extensibility of the protocol. Level 1 introduce only several metrics categories, which is acceptable for protocol extension. Level 2 metric only need one single metric, so it brings least burden to the protocol. Therefore, from the extensibility aspect, Level 2 and Level 1 metrics are suggested.¶
Regarding Stability, new Level 0 raw metrics may require new extension in protocol, which brings unstable format for protocol, therefore, this document does not recommend to standardize Level 0 metrics in protocol. Level 1 metrics request only few categories, and Level 2 Metric only introduce one metric to the protocol, so they are preferred from the stability aspect.¶
In conclusion, for computing-aware traffic steering, it is recommended to use the L2 metric due to its simplicity. If advanced scheduling is needed, L1 metric can be used. L2 metrics are the most comprehensive and dynamic, therefore transferring them to network devices is discouraged due to their high overhead.¶
Editor notes: this draft can be updated according to the discussion of metric definition in CATS WG.¶