Intent-based Meta-Scheduling in Programmable Networks: A Research Agenda

Nanjangud C. Narendra,Ronak Kanthaliya,Venkatareddy Akumalla

2024-12-05

Abstract:The emergence and growth of 5G and beyond 5G (B5G) networks has brought about the rise of so-called "programmable" networks, i.e., networks whose operational requirements are so stringent that they can only be met in an automated manner, with minimal/no human involvement. Any requirements on such a network would need to be formally specified via intents, which can represent user requirements in a formal yet understandable manner. Meeting the user requirements via intents would necessitate the rapid implementation of resource allocation and scheduling in the network. Also, given the expected size and geographical distribution of programmable networks, multiple resource scheduling implementations would need to be implemented at the same time. This would necessitate the use of a meta-scheduler that can coordinate the various schedulers and dynamically ensure optimal resource scheduling across the network. To that end, in this position paper, we propose a research agenda for modeling, implementation, and inclusion of intent-based dynamic meta-scheduling in programmable networks. Our research agenda will be built on active inference, a type of causal inference. Active inference provides some level of autonomy to each scheduler while the meta-scheduler takes care of overall intent fulfillment. Our research agenda will comprise a strawman architecture for meta-scheduling and a set of research questions that need to be addressed to make intent-dynamic meta-scheduling a reality.

Networking and Internet Architecture

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is: in programmable networks (especially 5G and B5G networks), how to achieve rapid and automated resource allocation and scheduling to meet strict low - latency and other Quality of Service (QoS) requirements. Specifically, the paper focuses on:

1. **Intent - driven meta - scheduling problems**:
- In large - scale, distributed, and complex programmable networks, how to express user needs through "intents" and ensure that these intents can be effectively decomposed and implemented.
- How to design and implement a meta - scheduler to coordinate conflicts among multiple schedulers and ensure optimal resource scheduling.

2. **Requirements for rapid resource allocation and scheduling**:
- With the development of 5G and B5G networks, resource scheduling needs to be completed within sub - millisecond levels to meet the strict latency requirements of 6G networks.
- Multiple user requests in large - scale networks need to be processed simultaneously, which may lead to conflicts among scheduling algorithms. Therefore, a meta - scheduler is required to coordinate these schedulers.

3. **Applications of causal inference and active inference**:
- Traditional machine learning methods have problems such as black - box nature and poor adaptability in resource scheduling. Therefore, causal inference based on causal models and active inference are introduced.
- Active inference can model how the brain predicts and evaluates sensory information to reduce long - term "surprises" and is applied to resource scheduling to improve the accuracy and efficiency of scheduling decisions.

### Main contributions of the paper

- **Proposing an intent - based meta - scheduling framework**: Combining the O - RAN architecture, a two - layer meta - scheduling framework including a meta - scheduling layer and a scheduling layer is proposed to achieve intent - driven resource management.
- **Introducing causal inference and active inference**: Using causal models and active inference to improve resource scheduling algorithms, making them more flexible and efficient.
- **Defining key research questions**: Several key research questions, such as intent decomposition methods, causal model discovery, and elasticity strategies, are proposed, providing directions for future research.

### Formula example

In the paper, some formulas are involved, such as the Free Energy formula in active inference:

\[
F[Q, o] = D_{KL}(Q(x) \| P(x|o, m)) + \mathcal{I}(o|m)
\]

where:
- \( F[Q, o] \) is the free energy,
- \( D_{KL}(Q(x) \| P(x|o, m)) \) is the Kullback - Leibler divergence, representing the difference between the approximate posterior probability \( Q(x) \) and the exact posterior probability \( P(x|o, m) \)
- \( \mathcal{I}(o|m) \) is the negative log - likelihood of the observed value \( o \) given the model \( m \)

This formula is used to measure the gap between the observed value and the model prediction and guide the scheduler to adjust its behavior to minimize this gap.

By solving these problems, the paper aims to provide a more intelligent and efficient resource management and scheduling mechanism for future programmable networks.