📖Introduction

Janction is the first Layer2 to provide verifiable, synergic and scalable Al service. Janction uses smart contracts to automate and dynamize machine learning and artificial intelligence. Key elements like Al models, GPU computing power, data feeding, and data labeling all are integrated into Janction for coprocessing.

We consider how to build such a collaborative system in a distributed trustless way. The following aspects need to be considered:

Resource Scheduling and Management: We need to consider the reasonable and safe allocation of resources such as arithmetic, data, storage, bandwidth, etc. in the system to ensure that the needs of various users and participants are met efficiently, reliably, and equitably, and most importantly, to address the arithmetic routing and cluster mixing section (Colocation).

Data Acquisition: Data acquisition is the process of extracting data from a variety of sources, we can consider a variety of ways to obtain user data from both off-chain and on-chain and give reasonable compensation to data providers.

Proof of Workload: In order to build a truly de-trusted computational network and provide financial incentives for participation, Janction must have a way to verify that the computational work of the AI is indeed performed as promised.

Gaming Mechanism: There are multiple roles in the system participating in the computational tasks, including data providers, data annotators, arithmetic providers, model providers, etc. We need to design a fair and reasonable gaming mechanism that makes it possible to both incentivize the roles in the system to complete the required tasks and maximize the benefits among the roles.

Privacy: At present, countries around the world are paying more and more attention to the protection of personal data privacy, and some countries have already legislated to manage it. We need to address the issues of data segregation, protection of data privacy and security while avoiding data silos during model training and model inference.

Parallelization: Parallelization of deep learning is the most important problem to be solved in the process of AI training, there are already many algorithms in the implementation of neural networks for multi-GPU training, mainly including data parallelism, tensor parallelism and pipeline parallelism, we design a reasonable parallelism, and we need to design a way to communicate in the process of distributed parallel computing.

Currently, distributed AI in academia and industry can only be realized up to model inference, and the architecture of Janction is mainly designed based on model inference.Janction will start from data acquisition, preprocessing, and GPU arithmetic market to gradually realize the AI collaborative Layer2 system.The focus of Janction will be mainly on the core design of Janction cluster GPU pool and its billing system. pool and its billing system core design.Janction Layer2 mainly addresses:

1. designing the way Janction accomplishes data acquisition and pre-processing from both on-chain and off-chain;

2. using microservice architecture to simplify the AI model deployment process by packaging algorithms, system and runtime optimizations and adding standard APIs, containerizing network resources to facilitate system scheduling and management, and providing an end-to-end computational power routing solution on top of that;

3. analyzed the scheme of efficiently scheduling GPU idle arithmetic using the mixing part algorithm and its rationality;

4. designing a GPU arithmetic market, constructing a role model and a gaming mechanism within the Janction system, and pricing the work of arithmetic providers using a workload correlation function with the participants and the VCG mechanism;