DNS-based Service Discovery for Computing-Aware Traffic Steering (CATS)

Introduction The Computing-Aware Traffic Steering (CATS) framework is designed to enable traffic steering that takes into account both network conditions and computing resource availability. A key requirement of this framework is providing a discovery and resolving method for the mapping of a service identifier to a speficic address . where computing resources are available. This document specifies how DNS-based Service Discovery (DNS-SD) can be extended and used to fulfull this requirement within the CATS framework. DNS-SD provides a standardized mechanism for service discovery using existing DNS infrastructure, making it well-suited for integration with the CATS architecture. The approach outlined in this document enables:

Publishing of computing service availability through DNS-SD
Discovery of appropriate CATS service instances based on service and resource requirements
Resolution of CATS service identifiers to specific network addresses.
Advertisement of CATS service capabilities and parameters using a dedicated, strucutured Resource Record.
Dynamic and secure updates of service availability and characteristics.

This document describes the necessary extensions to DNS-SD to support CATS-specific parameters and how the discovery mechanism integrates with other components of the CATS framework.

Requirements Language 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 when, and only when, they appear in all capitals, as shown here.

Terminology This document uses the terms defined in and .

Service Instance Names for CATS CATS service instances MUST be identified using DNS-SD service instance name following the format defined in RFC 6763 : Service Instance Name = <Instance>.<Service>.<Domain> where

the <Instance> portion is a user-friendly name for the instance
the <Service> portion indicates the name of a specific type of service
the <Domain> portion indicates the domain name where the service registered

Service Names As defined in RFC 6763 , the <Service> portion of a Service Instance Name consists of a pair of DNS labels _<Service-name>._<Proto> where _<Service-name> is a symbolic name of the desired service, and _<Proto> is the symbolic name of the desired transport protocol. For services using TCP, the second label is "_<tcp>", and for services using any transort protocol than TCP, the second label is _<udp> This document defines the following primary types of service for CATS services:

_cats-inference._tcp (for ML inference services)
_cats-storage._tcp (for storage services)
_cats-computing._tcp (for general computing services)

Service Instance Names Service instance names in CATS follow the DNS-SD convention: <instance-name>.service-name._tcp.<domain> Where:

<instance-name> is a user-friendly name for the service instance
<domain> is the DNS domain in which the service is registered

For example: edge-inference-1._cats-inference._tcp.example.com The instance name SHOULD be unique within the domain to avoid conflicts.

Service Parameters and CATSPARAM Records In RFC RFC 6763 requires that every DNS-SD service MUST have a TXT record in additon to its SRV record, with the same name, even if the service has no addtional data to store and the TXT record contains no more than a single zero Byte. However, it is widely recognized that using TXT records will introduce security issues such as being abused by malwares. To avoid the security and parsing issues associated with TXT records, this document defines a new DNS Resource Record type, CATSPARAM, for advertising CATS services parameters. The CATSPARAM RR is associated with a service instance name and contains a structured set of key-value paers representing the service instance's capabilities and state. The RDATA for a CATSPARAM record consists of a series of parameters blocks. Each block is a key = value pair, represented as a length-prefix string. This format allows for extensibility while maintaining a clear structure. The following parameters are defined:

"cpu": CPU capacity in normalized units (integer)
"mem": Memory capacity in MB (integer)
"lat": Expected processing latency in milliseconds (float)
"load": Current load level (0-100) as a percentage (integer)
"gpu": GPU availability and type (string)
"accel": Other accelerator availability and type (string)
"vers": Service version (string)
"caps": Capabilities as a comma-separated list (string)
"prio": Priority tier (integer, lower values indicate higher priority)
"cost": Relative cost metric (integer)
"avail": Availability status (0=offline, 1=online, 2=degraded)

SRV Records for Service Location For each service instance, an SRV record MUST be published according to RFC 2782 to enable clients to locate the service. The SRV record format for CATS services instance is: <instance-name>._cats._tcp.<domain> IN SRV <priority> <weight> <port> <target> Where:

<priority> represents the priority of the target host (lower values indicate higher priority)
<weight> is used for load balancing among targets with the same priority
<port> is the TCP port where the service is available
<target> is the hostname of the machine providing the service

For example: edge-inference-1._cats._tcp.example.com. SRV 0 5 8080 compute1.example.com.

Integration with CATS Framework

Relationship with CATS Control Plane The DNS-SD discovery mechanism integrates with the CATS control plane in the following ways:

The CATS control plane MAY act as a discovery client, querying for available computing services and maintaining a database of available resources.
The CATS control plane MAY facilitate service registration by providing interfaces and automation for DNS record management.
The CATS control plane MAY implement advanced selection algorithms that consider both the parameters advertised via DNS-SD and additional network and computing metrics.
For centralized deployments, the CATS control plane MAY provide a proxy service that mediates between clients and the DNS-SD infrastructure.

Service Parameter Advertising Service parameters advertised through DNS-SD TXT records provide inputs to the CATS framework's decision-making process for traffic steering. The computing service MUST ensure that advertised parameters accurately reflect the current state and capabilities of the computing resource. Parameters SHOULD be updated when significant changes in resource availability or characteristics occur. The CATS control plane MAY augment the DNS-SD parameters with additional information from other sources when making steering decisions.

Dynamic Updates Computing services MUST update their DNS-SD records when significant changes in availability or capabilities occur. These updates can be performed through:

Standard DNS Dynamic Update mechanisms
DNS Update Leases for time-limited registrations
Multicast DNS (mDNS) for local network scenarios

The frequency of updates SHOULD be balanced to reflect accurate information while avoiding excessive DNS traffic. Services SHOULD implement a dampening mechanism to avoid frequent updates for minor or transient changes. For highly dynamic parameters like current load, services MAY implement a threshold-based update policy, only updating the DNS records when the parameter crosses predefined thresholds.

Service Discovery Process and Protocol Flow The DNS-SD for CATS protocol flow can be shown in the following figure. | | | | | | | | 2. UPDATE_ RESPONSE(success) | | |<-----------------| | | | | | | | 3. CATS_REGESTER(svc_name, instance_name, | domain, host, port, params) | | |------------------------------------>| | | | | | | 4. CATS_REGISTER_RESPONSE(success) | | |<------------------------------------| | | | | | Phase 2: Discovery | | | | | | | | | 5. DNS_QUERY(type=PTR,svc_name) | | |<--------------------------------------| | | | | | | 6. DNS_RESPONSE(instance names) | | |-------------------------------------->| | | | | | | 7. DNS_QUERY(type=SRV+CATSPARAM,| | | instance name) | | |<--------------------------------------| | | | | | |8. DNS_RESPONSE(host, port, parameters)| | |-------------------------------------->| | | | | Phase 3: Selection and resolution | | | | | | | | 9. SERVICE_REQUEST(service type,| | | requirements)| | | |<-------------------| | | | | | | 10. SERVICE_SELECTED(hosts,port)| | | |------------------->| | | | | | | 11. DNS_LOOKUP(host) | | |<--------------------------------------| | | | | | | 12. DNS_RESPONSE(IP address) | | |-------------------------------------->| | | | | 13. CONNECT(IP address, port) | |<-------------------------------------------------------->| | ]]>

Registration Phase Registration can be performed using authenticated DNS update mechanisms or Dynamic DNS update protocols. For each CATS service:

Create a PTR record pointing from the service type to the service instance name
Create an SRV record specifying the host and port for the service
Create a CATSPARAM record containing the CATS-specific parameters
Send a DNS_UPDATE message to the DNS server to add or update these records to the zone file.
Register to the C-SMA with service name, instance name and parameters.

The DNS_UPDATE message MUST be authenticated using TSIG or SIG(0) as described in Section 9. To prevent name collision, the update transaction SHOUDLD use preconditions. For a new service, the update MUST assert that the PTR, SRV and CATSPARAM record names do not already exist (NXDOMAIN precondition). The CATS control plane MAY facilitate this registration process through an appropriate management interface.

Discovery Phase Clients requesting CATS services initiate the discovery process by querying for PTR records matching the appropriate service type. For example: _cats-inference._tcp.<domain> The response includes the list of matching service instance names. The client then queries for the SRV and CATSPARAM records associated with each service instance to obtain location information and service parameters.

Selection and Resolution Phase After obtaining the list of available services and their parameters, the client or the CATS control plane performs service selection based on the application requirements and the advertised parameters. The selection process MAY consider:

resource capabilities (CPU, memory, GPU, etc.)
Current load and availability
Expected latency and performance metrics
Priority and cost considerations
Specific capabilities required by the application

Once a suitable service is selected, the client resolves the hostname from the SRV record to an IP address using standard DNS A or AAAA queries, and establishes a connection to the service using the specified port.

Deregistration Phase When a service instance is terminated or becomes permanenetly unavailable, is MUST be deregistered. This ensures the clients no longer discover ot attempt to connect to the unavailable services. The deregistration is performed by sending an authenticated DNS_UPDATE message that deletes all records associated with the dervice isntance (PTR, SRV, and CATSPARAM). The delete request MUST be authenticated with the same credentials used for registration to prove ownership.

Implementation Considerations

Multicast DNS Considerations In local network environments, Multicast DNS (mDNS) MAY be used in conjunction with DNS-SD to provide service discovery without requiring a centralized DNS server. When using mDNS, CATS services SHOULD:

Respond to mDNS queries for their service type
Advertise their presence periodically as specified in RFC 6762
Implement proper conflict resolution mechanisms
Consider the scope and scale of the deployment, as mDNS is primarily designed for local network use

DNS-SD/DNS Integration For larger-scale deployments across multiple networks, traditional unicast DNS infrastructure is RECOMMENDED. In these scenarios:

CATS services SHOULD be registered in appropriate DNS zones
DNS infrastructure SHOULD support DNS Dynamic Updates
DNS servers SHOULD be configured to allow updates from authorized CATS components
Consider using DNS Update Leases for time-limited registrations
Implement appropriate caching policies for DNS records

Performance Considerations Implementers SHOULD consider the following performance aspects:

DNS query volume: In large deployments with many clients, implement appropriate caching and consolidation of discovery requests.
Update frequency: Balance the need for accurate information with the overhead of frequent DNS updates.
Scalability: For very large deployments, consider hierarchical discovery approaches or specialized discovery proxies.

IANA Considerations This memo includes no request to IANA.

Security Considerations The use of DNS-SD for CATS service discovery introduces several security considerations:

Authentication: DNS updates for service registration SHOULD be authenticated to prevent unauthorized registration of services. DNS Security Extensions (DNSSEC) SHOULD be implemented to provide authentication of DNS data.
Information disclosure: Service parameters may reveal sensitive information about computing capabilities and deployment details. Consider the privacy implications of the parameters being advertised.
Denial of Service: Large-scale DNS-SD queries could potentially be used for denial-of-service attacks. Implement rate limiting and monitoring for unusual query patterns.
Spoofing: Without DNSSEC, DNS responses could potentially be spoofed, leading to service misdirection. DNSSEC validation SHOULD be enabled for DNS-SD queries.
Data integrity: Ensure that computing parameter updates go through proper validation to prevent advertising incorrect capabilities, which could lead to suboptimal traffic steering decisions.