MRAM-Centric Storage System for Low-Latency Data Persistence
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional storage systems face limitations in providing low latency, data persistency, high endurance, and power efficiency required for high-performance processing, particularly due to the constraints of memory-storage hierarchies involving 3D XPoint DIMMs and Optane SSDs, which result in increased write amplification, higher access latency, lower throughput, and lower endurance compared to DRAM.
Innovation Solution
A non-volatile storage system utilizing MRAM as the main memory, NOR for low-latency reads, NAND for data reliability and page swapping, and HDD for storing cold data classified by access frequency, allowing transparent data movement between these media without relying on DRAM or logical storage drives, thereby optimizing the usage of MRAM and mitigating data retention issues.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a memory-storage hierarchy with additional tiers (3DXP DIMMs, Optane SSDs) is used, then data persistency is improved, but access latency increases and throughput decreases compared to DRAM
Solution Approach 1:
The patent merges memory and storage functions into a unified non-volatile memory-centric system. MRAM serves as both main memory and persistent storage, eliminating the traditional memory-storage hierarchy. This integration allows the system to achieve both low-latency access (memory-like) and data persistency (storage-like) simultaneously, resolving the contradiction between speed and reliability.
Solution Approach 2:
MRAM is designed to perform multiple functions: it operates as volatile memory for high-speed access, as non-volatile storage for data persistency, and as a unified addressable space for both. This multi-functionality allows the same physical medium to satisfy both the low-latency requirements of memory and the reliability requirements of storage, eliminating the trade-off between access speed and data persistency.
2Loss of time
If MRAM is used as main memory, then access latency is reduced, but cost increases compared to DRAM
Solution Approach 1:
The system segments the addressable space into different regions with different retention requirements. Frequently accessed data (hot data) is placed in MRAM for low-latency access, while less frequently accessed data (cold data) is migrated to lower-cost non-volatile storage media. This segmentation allows the system to use expensive MRAM only where its low-latency benefits are most valuable, reducing overall system cost while maintaining performance.
Solution Approach 2:
The system dynamically changes the retention parameter of data by migrating it between storage media based on access patterns. Hot data maintains high retention in MRAM, while cold data is transferred to lower-cost media with adequate but lower retention characteristics. This parameter change allows the system to optimize the cost-performance trade-off by matching data retention requirements with appropriate storage media.
3Reliability
If data is copied between multiple storage tiers, then data persistency is improved, but write amplification and power consumption increase
Solution Approach 1:
The system implements self-service data management through automated tiering and caching mechanisms. The storage manager monitors access patterns and automatically migrates data between MRAM and non-volatile storage media based on retention requirements and access frequency. This self-service approach eliminates the need for manual data copying and optimization, reducing unnecessary write amplification and power consumption while maintaining data persistency.
Solution Approach 2:
The system performs preliminary data migration and caching actions based on predicted access patterns. By proactively moving data to appropriate storage tiers before access occurs, the system reduces the need for frequent data copying and repositioning. This preliminary action minimizes write amplification and associated power consumption while ensuring data is in the optimal location when accessed.
4Reliability
If 3DXP DIMM is used, then data persistency is improved, but endurance is reduced compared to DRAM DIMM
Solution Approach 1:
The system uses MRAM as a disposable cache layer for frequently accessed data that requires high endurance. Data that can tolerate lower endurance is stored in lower-cost non-volatile storage media. This approach allows the system to protect against data loss by maintaining copies in both high-endurance (MRAM) and low-endurance (non-volatile media) locations, using the cheaper medium for bulk storage and the more durable medium for protective caching.
Data Source
AI summary
A system is provided to receive a first request to write data to a non-volatile storage system, which comprises an MRAM, a NAND, and an HDD. The system allocates a first physical address in the MRAM and writes the data to the MRAM at the MRAM first physical address. In response to determining that the data in the MRAM is not accessed within a first predetermined time period, the system copies the data from the MRAM to the NAND at a NAND physical page address and maps a logical page index associated with the data to the NAND physical page address. In response to determining that the data in the NAND is not accessed within a second predetermined time period, the system copies the data from the NAND to the HDD based on an HDD physical address and maps the NAND physical page address to the HDD physical address.