Loading...

Improving Cache Performance of Data Storage Systems Using Machine Learning

Ebrahimi, Shahriar | 2017

379 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 50261 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Asadi, Hossein
  7. Abstract:
  8. Emerging Solid State Drives (SSDs) have performance advantages over traditional Hard Disk Drives (HDDs). Higher price per capacity and limited lifetime, however, prevents enterprise data centers to entirely replace HDD-based storage subsystems with SSDs. Thus,SSD-based caching has been widely employed in data centers to benefit from higher performance of SSDs while minimizing overall cost. Input/Output (I/O) workloads exhibit unpredictable and highly random behavior which makes conventional algorithms such as Least Recently Used (LRU) not able to provide high hit ratio as they employ linear localities.In addition to poor performance, such algorithms also shorten SSD lifetime with unnecessary cache replacements. In this research, we propose the first reconfigurable SSDbased cache architecture using Recurrent Neural Networks (RNNs) to characterize ongoing workloads and can optimize itself towards higher cache performance while increasing SSD lifetime. Proposed method consists of an offline and an online phase. In the offline phase,we try to learn various workloads and predict their behavior. In the second phase, collected information gets used to identify performance critical data pages to be cached. Experimental results show that proposed method can characterize workloads with an accuracy up to 94.6% for SNIA I/O workloads. The proposed method can perform similarly to optimal cache algorithm by an accuracy of 95% on average and outperforms previous SSD caching architectures by having up to 7x and 10x higher hit ratio and endurance, respectively
  9. Keywords:
  10. Machine Learning ; Cache Memory ; Characterization ; Neural Networks ; Data Storage ; Solid State Disk Drive

 Digital Object List