What is DEKUBE?FULL STACK AI NETWORK SOLUTION FOR TRAINING, FINE-TUNING, INFERENCE.
How does DEKUBE work?Hover on the diagram to read more about key elements.
6
- About DEKUBE Network
- DEKUBE is a distributed AI computing network.
- The decentralized, secure, and high TPS features of the P2P network provide the foundation for DEKUBE to achieve large-scale distributed device collaboration.
- Built on Kubernetes (K8s) technology, it constructs containers and schedules computing resources in distributed GPU devices to ensure task consistency, efficient resource utilization, and flexibility.
- The Network Gateway provides a secure and efficient communication infrastructure between DEKUBE nodes, ensuring privacy and data exchange speed.
- About Container Service
- The container service achieves environmental consistency, providing an encapsulated, independent runtime environment for applications and their dependencies. It accomplishes the goal of consistent task execution on distributed GPU devices of different models and environments.
- The Container Engine is responsible for creating, running, and managing the runtime environment of containers. An efficient containerized environment supports the execution and management of large-scale distributed tasks.
- GPU Access Dispatch Service efficiently manages and distributes GPU resources to support deep learning and other computationally intensive tasks in DEKUBE Node.
- About Tasks Launching Service
- The Task Launch Service is responsible for parsing, calculating, launching, and managing the lifecycle of tasks.
- It parses task information, calculates the resource cost required to complete the task, ensuring that the task can be accomplished on the node.
- It monitors and records the execution status, resource consumption, and performance of tasks, sending and recording task execution results on the P2P network.
- It is responsible for scheduling tasks, and if the resources are full or the task execution fails, it will schedule the task to other available nodes.
- About Pod Control Agent
- The Pod Control Agent is responsible for managing and scheduling containers, including the startup of containers, task orchestration, and various stages of the container lifecycle. It helps achieve efficient utilization of computational resources on each node in DEKUBE.
- The Pod Launch Scheduler is responsible for scheduling and launching containers (Pods). In the event of a computational node failure or task execution failure, it is responsible for rescheduling tasks to other available nodes to ensure the reliable execution of tasks.
- The Tasks Orchestration assigns different tasks to available computational resources to maximize the resource utilization of each node.
- The Pod Monitor monitors and collects information on the resource usage, health status, and progress of tasks being executed by Pods.
- About Pods
- Pods are the basic units for containerized resource scheduling in DEKUBE nodes, and all tasks are segmented and assigned to Pods for execution.
- Pods can easily scale up and down, automatically adjusting based on changes in workload. This contributes to more efficient utilization of computational resources, enhancing the performance and efficiency of DEKUBE.
- Pods provide a means of isolating tasks and their dependencies, ensuring that tasks within containers can be executed independently in separate environments, avoiding conflicts.
- Containers within the same Pod can access the same local storage volume, facilitating data sharing for collaborative tasks.
- About Blockchain Interaction Message Queue
- The interaction between DEKUBE nodes and the P2P network is processed through the Blockchain Interaction Message Queue.
- Transactions include tasks acquisition, feedback on task records and results, fund accounting, and more.
- The message queue provides a certain level of security and reliability, ensuring that crucial messages or tasks are not lost in communication between the system and the blockchain.
- About P2P Network
- P2P network is a pioneering high-speed decentralized secure blockchain leading the industry with a TPS of 12,000. Featuring a proprietary high-speed peer-to-peer network layer surpassing traditional counterparts, it caters to the demanding network transmission needs of AI training.
- It will provide DEKUBE with decentralized interaction between all participants and trust between them.
- P2P network participates in specific transactions of DEKUBE, including the complete lifecycle of task recording, settlement, and maintaining the network status of all nodes.
- About Network Gateway
- The network gateway is a blockchain-enhanced peer-to-peer network layer and is the software core of DEKUBE.
- It builds a secure and efficient communication infrastructure, ensuring the security of communication, maintaining node network status, and providing a solid foundation for data transmission and task coordination support in the DEKUBE.
- It provides efficient peer-to-peer communication, ensuring the rapid and reliable exchange of information between nodes, with an average speed 6 times higher than the traditional blockchain's peer-to-peer network layer.
- It utilizes blockchain technology to achieve secure data transmission, ensuring the integrity and confidentiality of information in the network.
- About Client For GPU Provider
- GPU providers can install with one click through the client, completing the local computing resource environment deployment. Through the client, users can easily manage local computing resources, connect to the cluster, and claim rewards.
- About Artifactory Server
- Artifactory Server is a repository management system used for managing and storing software builds, artifacts, and dependencies.
- In DEKUBE, different tasks, services, or components may depend on shared libraries, frameworks, or other dependencies. Artifactory can serve as a central repository, storing these dependencies, ensuring their consistency and availability across various nodes in the network.
- Through its caching mechanism, Artifactory can reduce direct dependencies on external resources, thereby improving the speed of builds.
