Providing Instance Affinity in Dyncast

Status of This Memo

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 26 August 2021.¶

Copyright Notice

Copyright (c) 2021 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.¶

Table of Contents

1. Introduction

Dyncast support in the network provides a client with a fresh optimal path to a service provider instance, where optimality includes both path and service provider characteristics. As a service invocation usually takes more than one packet, dyncast needs to provide instance affinity for each service invocation. Naive implementations of instance affinity require per-application, per service-invocation state in the network.¶

The present short document defines a way to provide instance affinity that does not require, but also does not rule out per-application state.¶

[I-D.liu-dyncast-ps-usecases] lists use cases of dyncast. The present document does not discuss the specifics of how the network provides dyncast, such as the way service instance metrics enter path computations.¶

2. Terminology

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 the terminology of [I-D.liu-dyncast-ps-usecases], in particular Service and Service Instance (the latter often abbreviated to "Instance"). It also defines the following terms:¶

Client:

The system that requests a service.¶

Service invocation:

A single transaction between client and a service instance. The client is interested in talking to the same service instance throughout one service invocation. Subsequent and parallel service invocations can use different service instances without a problem and therefore do not require affinity.¶

Instance Affinity:

The ability of the network to send all the packets of a service invocation to the same service instance. (Note that this doesn't necessarily imply path affinity -- the client does not care about the path, only about getting to the same service instance.)¶

Service period:

The temporal granularity (rhythm) in which the network updates the optimal paths it provides for a service.¶

Service stretch:

The maximum amount of time that the network plans to provide instance affinity for a service invocation.¶

3. Assumptions

This document makes a number of assumptions, some of which are fundamental to its technical approach, but some of which are only required for the exposition chosen in this document. A future version of this document will clearly separate these two kinds of assumptions.¶

Due to experience with overly eager load-based updates to routing metrics, we assume that metrics will be updated on the scale of tens of seconds. To simplify exposition we therefore set the service period to 10 seconds (assumptions of this kind are intended to be possible without loss of generality, but should not be wildly off).¶

We assume the affinity processing for the entire network will be on a rhythm that is consistent with the service period. Updates take effect at the start of a new service period. The entire network is loosely synchronized on this rhythm. The clients are also aware of this rhythm.¶

We assume the service stretch will be quite limited, on the order of (a generous) five minutes or less. As a result, any service invocation covers less than 32 service periods. Services that do need longer service stretches will need to renew the service invocation regularly (by checking whether the service instance has changed upon such a renewal, any handover effort needed can be minimized).¶

Service identifiers take the form of IPv6 addresses, or more precisely, IPv6 prefixes. The client is able to complete the prefix with application information. (In a pinch, the client can obtain a complete current address via DNS lookup.)¶

4. Objectives

Dyncast needs to provide instance affinity. The present document outlines how to achieve this without creating per application, or worse, per invocation state in the network.¶

The network does not provide any signaling to the clients beyond what is expected in an IPv6 environment.¶

In summary, the objective of this draft is to define a stable client interface to the instance affinity mechanism (and to motivate why this interface is useful). This interface is designed to remain stable even while the network support for this mechanism is evolving.¶

5. Approach

We number the service periods with a cyclic numbering system that wraps around about every two service stretches. The network and the clients are aware of the current service period number; the synchronization requirement between them is that clients typically aren't ahead of the network.¶

When starting a new service invocation, the client builds an IPv6 address out of the service identifier and its view of the current service period number (or it obtains this address using a DNS lookup), essentially filling in 6 bits (for the numbers assumed here). Service requests and the resulting communication within the invocation are addressed to this current address. The client stores the current address with the service invocation when initializing it; it is not ever updated for this invocation.¶

The network keeps its path optimization state relative to (or indexed by) the current period number. Routing updates can be processed at any time but do not lead to an update of the path optimization state for any service period. The result is that the path chosen after a routing update may no longer be optimal, but that instance affinity is kept. For each service, a pointer for the best service instance is kept for the current and the last 32 service periods.¶

6. Discussion

The approach presented provides instance affinity without requiring per application or per invocation state in the network. It does require up to 32 copies of what are essentially host routes per service instance. The state scales with the number of service instances, and not with the number of clients.¶

The approach is based on IPv6. It can be made to work in an IPv4 network, if there are plentiful IPv4 addresses available (see also Section 8).¶

7. Details

The service period number could simply be inserted in the service identifier, or more complex computation could be performed to make the current addresses generated this way stand out in a forwarding engine.¶

Naïve clients will start a service invocation with a DNS lookup. This allows the insertion of the period number to be performed in a specialized DNS server for the service. Of course, this requires short time to live (TTL) values and clients that do not on their own cache the look up results.¶

So the preferred variant is for the client to be aware of the current service period number and to do the insertion by itself on each new service invocation.¶

8. Legacy IP Considerations

To make this work with IPv4 addresses as service identifiers, we would need 6 bits that can be varied over time. This is likely too expensive for many applications. An alternative approach is to use the port number for the 6 bits. This would mean that the network would need to look up paths both on destination IP address and destination port number (48-bit addressing). For IPv4, this should be good enough.¶

9. Security Considerations

TBD¶

10. IANA Considerations

No IANA action is required for this concept draft.¶

Author's Address