Distributed cache processing method, device, storage medium and program product

By randomly discarding the location indexes of cache misses in a distributed shared cache pool and retrieving sample data from cloud storage, the data I/O bottleneck problem in deep learning model training is solved, improving cache hit rate and training efficiency.

CN122152195APending Publication Date: 2026-06-05ALIBABA CLOUD COMPUTING CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
ALIBABA CLOUD COMPUTING CO LTD
Filing Date
2024-12-03
Publication Date
2026-06-05

AI Technical Summary

Technical Problem

During the training of deep learning models, data I/O time overhead becomes a bottleneck, especially in distributed training. Due to the network bandwidth limitations of cloud storage, the data transmission speed is lower than the local retrieval speed. Existing caching technologies have insufficient cache hit rate under limited resources and cannot effectively reduce data I/O time overhead.

Method used

By randomly discarding some cache-missing location indexes in a distributed shared cache pool and obtaining corresponding sample data from cloud storage, combined with a first-in-first-out (FIFO) mechanism, the sample data is stored in the cache pool, thereby improving the cache hit rate and reducing data I/O time overhead.

Benefits of technology

It improved cache hit rate, reduced the performance overhead of loading sample data from cloud storage, and improved the efficiency of deep learning model training.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122152195A_ABST
    Figure CN122152195A_ABST
Patent Text Reader

Abstract

Embodiments of the present application provide a kind of distributed cache processing method, equipment, storage medium and program product, the method comprises: computing node determines the second quantity of position indexes and the third quantity of position indexes in remaining from the first quantity of position indexes required for obtaining this iteration according to the respective position index of cached sample data, to set probability random discard the third quantity of position indexes, to determine the position index and discarded position index retained therein. The sample data corresponding to the position index retained is obtained from cloud storage space. The sample data corresponding to the second quantity of position indexes is obtained from cached sample data. The sample data of discarded position index quantity is sampled from cached sample data to replace the sample data corresponding to discarded position index. More sample data is obtained from cache pool by the present scheme, and the cache hit rate is improved.
Need to check novelty before this filing date? Find Prior Art