Solution Description

Basic Concepts

• Ascend AI processor-based affinity
  • Principles: Affinity relies on the Ascend AI processor's interconnection topology and processing logic to maximize utilization efficiency.
  • Affinity scheduling policy: Based on the affinity rule, Volcano selects the scheduling logic of a specific Ascend AI processor. By adhering to the affinity scheduling policy and principles, resources can be allocated optimally.
• Node-based affinity
  • Switch affinity scheduling: Based on the networking configuration and parameter plane network configuration of switch nodes, optimal node utilization can be achieved.
  • Affinity scheduling of logical SuperPoDs: A physical SuperPoD in a cluster can be divided into logical SuperPoDs based on the division policy to achieve optimal utilization.

Ascend AI Processor-based Affinity

This document describes affinity principles of Ascend AI processors of Atlas training product, Atlas 200T A2 Box16 heterogeneous subrack, and A200T A3 Box8 SuperPoD Server, which are formulated based on the Volcano scheduling rules.

Node-based Affinity

This document also describes node-based affinity principles of Atlas training product and Atlas A2 training productAtlas 900 A3 SuperPoD, that is, node scheduling rules of switches. It further provides guidance on how to select proper switch nodes within the spine-leaf network architecture.