Tierra: multi-tiered arrays and recency-aware hot data decision.

Area of Science:

• Computer Science
• Data Management
• System Optimization

Background:

• Hot data identification is crucial for computing systems, especially with non-volatile memory (NVM).
• Existing methods using bit-array counters or multiple bloom filters struggle with accuracy, overhead, and memory due to ineffective data structures.
• Limitations in capturing both data recency and frequency information hinder accurate hot data identification.

Purpose of the Study:

• To propose a novel hot data identification scheme, Tierra, that overcomes the limitations of existing approaches.
• To introduce a new data structure based on asymmetric multilevel arrays for improved performance and accuracy.
• To reduce computational overhead and enhance identification accuracy through a recency-aware request screening mechanism.

Main Methods:

• Developed Tierra, a novel hot data identification scheme utilizing asymmetric multilevel arrays.
• Implemented a recency-aware request screening mechanism employing an enhanced stack distance approximation algorithm.
• Evaluated Tierra's performance and accuracy using diverse real-world workloads.

Main Results:

• The asymmetric multilevel arrays in Tierra reduce internal data movement by 3.1×.
• Tierra significantly reduces computational overhead while improving identification accuracy.
• Comprehensive evaluations demonstrated an average true identification rate of 99.4% for Tierra.

Conclusions:

• Tierra offers a highly accurate and efficient solution for hot data identification.
• The novel data structure and screening mechanism effectively address the limitations of previous methods.
• Tierra provides a promising approach for optimizing memory hierarchy, database, and storage systems.