Machine Learning Computational Optimization Methods and Platforms

Abstract

A machine learning calculation optimization method and platform are disclosed. The method includes: dividing a machine learning computation graph into a data work subgraph composed of upstream nodes of stateful nodes and a training work subgraph composed of stateful nodes and their downstream nodes; On the side, add a data sending node to the data work subgraph, and add a data receiving node to the training work subgraph. As a result, the computing graph is divided based on the presence or absence of nodes and the communication nodes are inserted. The present invention can decouple the data and training work in the same task, so as to dynamically allocate general computing resources that participate in data work at runtime, and solve problems such as GPU The problem of reducing the efficiency of deep learning tasks due to the provision of sufficient preprocessing data by dedicated computing units. Furthermore, by combining with the scheduler, general computing resource scheduling can be performed within the cluster, breaking the single-machine boundary and improving the overall hardware utilization efficiency of the platform.

Description

technical field [0001] The present disclosure relates to the field of machine learning, and in particular, to a method and platform for machine learning calculation optimization. Background technique [0002] Currently, data processing and training for deep learning tasks reside in the same piece of code, compiled together and run on the same machine. However, the ratio of general-purpose computing resources (for example, CPU) and special-purpose computing resources (for example, GPU, ASIC) required by different deep learning tasks is quite different. meet mission requirements. And with the improvement of the computing power of a single dedicated computing resource, the general-purpose computing resources usually equipped in the prior art cannot provide enough data for the dedicated computing resources, resulting in the operating efficiency of deep learning tasks caused by the mismatch of general-purpose and dedicated computing power. reduce. [0003] Therefore, there is ...

AI Technical Summary

Problems solved by technology

This scheme divides the computing graph based on the presence or absence of nodes and inserts communication nodes, so that the data work and training work in the same deep learning task are decoupled from each other, so that the general computing resources involved in data work can be dynamically allocated based on the efficiency of the runtime training work. Solve the problem of running deep learning tasks inefficiently due to the inability to provide enough pre-processed data to dedicated computing units such as GPUs

Images