Loading...

A Machine Learning Approach for Memory Resource Management in GPUs

Mohammadi Lak, Masoud | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57321 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Sarbazi Azad, Hamid
  7. Abstract:
  8. Nowadays, GPUs are used in applications such as graph computations and deep learning that require large data processing. These applications necessitate transferring a considerable amount of data between memory and computational units, which increases the average memory access latency. On the other hand, the poor performance of the L1 cache causes many reservation fails, resulting in increased energy consumption and reduced performance. In this thesis, we propose Smart Cache, a new smart cache management in GPUs. This approach models the cache replacement problem as a sequence labeling problem. Then, a machine learning model that combines support vector machine with one-hot encoding and k-sparse features is trained based on past accesses to the L1 cache of a streaming multi-processor to predict a priority score for evicting data from the L1 caches of streaming multi-processors. A replacement policy based on re-reference interval prediction is also employed to improve data eviction management. Evaluation results show that the Smart Cache improves the cache hit rate by 12%, performance by 24%, power by 3%, and energy consumption by 15% over LRU. The storage overhead of this approach is also negligible and less than 1% of the cache capacity of streaming multi-processors
  9. Keywords:
  10. Cache Memory ; Graphics Procssing Unit (GPU) ; Machine Learning ; Reservation Fails ; Energy Consumption Reduction

 Digital Object List

 Bookmark

...see more