Published in


NuNet architecture and functionality

NuNet is a global economy of decentralized computing. It allows computational runtimes, open-source and proprietary, to run in parallel and be combined, ready to be executed on any device that’s offering its hardware resources to the network. These runtimes can draw on data from various sources within the network, and any data provider can offer their data to be used by the network, and get crypto payments for it.

This post delves into how NuNet achieves this, taking a closer look at the components necessary to build this environment, and at how they fit together to enable emergent properties.

To give a brief overview, NuNet consists of –

1. Tokenomic incentives, a key part of the infrastructure that nurture a multi-sided marketplace in which computational resources are being actively exchanged.

2. Adapters: lightweight executables that run on each device on-boarded onto NuNet, connecting every device to control and data planes of the platform. The adapters track the computational resources and data posted to the network, search and match supplied resources to the demands of consumers, construct computational workflows and calculate the cost and time involved in executing them.

3. An open-ended collection of pluggable, open and evolving APIs (an API of APIs) that the adapters use to interact with the platform and other machines on the data and control planes.

Tokenomic layer

The tokenomics of NuNet is an essential part of the system; it attracts the community members into the system to participate, and motivates them to expand the network. Token payments make the link between a compute consumer in need of computational resources, and a community member with available resources. In this way, the blockchain can be thought of as another layer of the decentralized computing network.

At first, NuNet will support Ethereum, with full support planned for Cardano in the very near future. NuNet can use any blockchain with the necessary smart contract and micropayment features, and in later stages will provide the capacity for network operation agents to choose a smart contract and transaction platform from a number of options and framework extensions.

Our tokenomic mechanism will be discussed in detail in later blog posts.

Adapters and APIs

Adapters are executables that run on every device connected to NuNet and carry out NuNet functions, turning a device into part of the global decentralized computing economy. The adapter performs the functions necessary to integrate the device with NuNet, including measuring the compute resources available on the host device, tracking performance in real time, establishing communication with other machines on the network, and more.

The API of APIs is the shared language that adapters use to request resources, fulfill requests, and share information about network resources and topology. Every device on the network runs the NuNet adapter, but uses different APIs depending on its role in the network. For example, compute provider nodes use the Resource Description API to advertise their services, and the Service Ingestion API to onboard incoming tasks. Consumer oriented APIs are different: they will use the Task Description API to request specific services and resources, and the Process Validation API to check that their work is done.

The set of APIs includes –

a. Resource Description API – Compute providers use the Resource Description API to broadcast the nature and quantity of their resources to the network;

b. Data Description API – Data providers use the Data Description API to broadcast their data offerings (metadata, pricing, updates) to the network;

c. Data Ingestion API – Data consumers use the Data Ingestion API to receive data from data provider nodes. The transfer of data is tracked in a verifiable way so that token payments can be settled, considering privacy, provenance and persistence requirements;

d. Resource Discovery API – Compute consumers and AI algorithms use the Resource Discovery API to find resource providers that meet their computing needs;

e. Reputation API – Nodes use the Reputation API to rate compute suppliers, so that other customers can be assured they will perform computation faithfully. See the section on ‘Verification, validation, and reputation’ below.

f. Process Validation API – Customers and nodes that specialise in assessing reputation use the Process Validation API to verify that resource and data providers have fulfilled the requests made of them. See the section on ‘Verification, validation, and reputation’ below.

g. Workflow Aggregation API – Nodes acting as meta-agents use the Workflow Aggregation API to combine simple AI services into complex workflows resulting in hardware and service meshes within NuNet platform. See the section on ‘Workflow Composition’ below.

h. Topology API – The Telemetry API allows nodes to gain global awareness of the network. This helps meta-agents route around damage and dynamically design optimal workflows. This API also helps reputation-tracking systems to verify that data and compute has been sent through the network.

i. Tokenomics API -- implements the ability for each hardware device onboarded on NuNet to send and receive data on blockchain, sign transactions and cross-chain functionality of NuNet token; Tokenomics API acts as a wrapper for currently enabled or planned blockchain integrations (e.g. Ethereum, Cardano, and others);

j. Provenance API – Adapters use the Provenance API to verify the origin of data from data providers on NuNet, and of compute sourced from compute providers.

In short, a resource provider on NuNet runs an instance of the adapter, which stores information about its resources (processing power, memory, bandwidth, and more) and advertises this information to the network via the Resource Description API. This information is passed around in a decentralized way, rather than to a unified inventory, so that no one meta-agent oversees the whole network or stores a centralized book of the agents and offerings.

The decentralized control plane

Complex computing infrastructures manage the control plan using centralized ‘control planes’. Control planes are elements in the network architecture that configure the routes of data flowing through the network, balance load, and optimise the distributed mesh network from a single point of control.

NuNet is a decentralized hardware and software mesh architecture, and NuNet’s adapters form a decentralized control plane. The decentralized control plane allows agents to spin up ad hoc computational workflows, bringing together multiple software containers and independent hardware devices. It is decentralized, because there is no single point of control, as in conventional service mesh control planes implemented on the cloud. The platform will be able to design and execute computational and data-intensive services entirely autonomously, allowing computational agents to settle payments between themselves automatically on the blockchain.

In this way, NuNet’s architecture maximizes the mobility of computational processes. It enables them to move seamlessly between devices of the network. Workflows can dynamically optimize their own processes – bringing data closer to processes, or processes closer to data. Imagine a financial analysis AI that crunches data from the Tokyo stock exchange during its opening hours, and the New York Stock Exchange when it opens some hours later. At one time of day, it will be more efficient to compute this data in Asia, and later to compute it in North America; NuNet APIs allow dynamic reconfigurations like this.

The dynamic reconfiguration is possible because NuNet adapters are context aware. A key function of the Topology API is to provide each node with context awareness i.e. let each node know its location in the network and its proximity to other nodes. (This may include parameters which are agent-specific, and therefore could not be centrally managed.) This decentralized topology, with each node being aware of its own local environment, allows us to optimize computation in new ways: computational workflows can be mapped, and designed to minimizes the number of hops on the network, or minimize latency (transmission time) to execute the workflow. Or context awareness can be used by agents to query local network topologies in order to optimize for the cost (in tokens) of an operation.

Eventually, this architecture will enable users to search and discover computational processes automatically or semi-automatically, optimizing for speed or cost, or a mixture of both.

Note that NuNet workflows can operate across IoT, mobile, edge devices, private installations and even cloud computing vendors or more. For example, half your process might run on an external compute provider, and half on your mobile device, while all being described by the same computational workflow. The user can, however, specify restraints on what machines run his processes, such as constraints based on data security/privacy. This is hardly possible in the current cloud computing environments.

Workflow Composition

Some agents will be ‘meta-agents’, which is to say they will specialise in curating other agents’ computational services. These meta-agents will hold only the declarative description of a workflow (technically – a program graph, or a DAG), not its full representation in code, and will design the execution of workflows for other agents based on information they receive from APIs. This information will consist of the identities of agents in workflow, their inputs, their outputs, their cost, location and data offered, and information about scheduling.

In this way, the design of complex workflows will involve other agents. This federated structure in principle enables the composition of workflows of infinite scale and logical complexity on NuNet.

Meta-agents will be able to design workflows involving different levels of meta-agents, sub-meta-agents, and sub-sub-meta-agents, all the way down to base agent services. As each of these elements change their costs, the meta-agent can recalculate the overall cost on the fly. These workflows may be designed by a human operator, but more likely will be designed at least partly by automated systems, and progressively will become more autonomous, until minimal human intervention is required. Meta-agents could be SingularityNET AI Services that use SingularityNET’s AI capabilities to design optimal workflows.

Once the optimal way to execute the workflow is calculated, it can run independently, after verifying that the token-costs of all computational agents are covered.

The network-wide telemetry information that the meta-agent receives from the Telemetry API will facilitate self-learning and healing capabilities of the network, which themselves will be implemented as workflows running on NuNet platform. They will allow nodes in the network to detect damage to a remote part of the network, and route around it. This minimises the potential impact that bad actors may have on network performance, and allows processes to sail smoothly through the open network.

Verification, validation, and reputation

NuNet will provide tools for nodes to check that computational work was completed correctly. (Otherwise bad actors could lie, saying they have performed compute that they have not, to claim rewards.) Like everything, this validation is done in a decentralized way; NuNet as a whole does not provide an official check mark to guarantee the correctness of a computation. Instead, NuNet includes the Validation API, which will let agents validate computational processes running on their metal, plus the Telemetry API and Reputation API to gossip validation information around the network.

Over time, machines and compute providers that faithfully execute every task given to them will gain a good reputation among the swarm of other computers. As a good reputation will attract more customers, reputation is an economic incentive that motivates compute providers to faithfully execute their work. Blockchain integration and the Tokenomics API will enable tokenomic and reputation incentives that reward good behavior and punish bad behavior.

In the future, NuNet will provide an API for integrating third party formal verification tools (such as zkSNARKs, tools developed by SingularityNET, or other open source protocols) as additional options for assuring customers that compute has been carried out correctly.

Agents running on NuNet will be able access network statistics and information about their own and other agents’ capabilities in a secure and permissioned manner. For example, if computational agents have machine learning capabilities, they will be able to use this to form insights about the credibility, efficiency and security of other constituents of the network. Some agents may specialize in this: analyzing other agents’ reputations and rating their performance, and then providing this information to other agents in exchange for tokens or data. The agents with this functionality will together form NuNet’s decentralized reputation system, utilizing the Reputation API.

This unique architecture is totally decentralized and constitutes a multi-sided marketplace. NuNet use tokenomic incentives to attract a crowd of agents of various kinds – buyers, sellers, meta-agents, reputation-rankers, and more – who will form a global computing hardware and service mesh in which each workflow is intelligently designed and dynamically re-designed as it flows freely across an open network of borderless computational power.

About NuNet: NuNet lets anyone share and monetize their computing resources, turning cloud computing power from a centralized service into an open protocol powered by blockchain. Join our community via:



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store