Cluster File System Metadata Storage via DRAM SSD HDD Tiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional metadata storage systems based on DRAM + HDD architecture face performance bottlenecks due to increasing metadata volume and I/O demands, while SSD integration faces issues with small granularity random I/O modes that reduce SSD lifespan and increase costs, especially in cluster file systems.
Innovation Solution
A metadata storage system utilizing a DRAM + SSD + HDD three-tier architecture with sequential write-back and data migration based on access heat, employing Bloom Filters and hash tables for efficient metadata indexing and storage, and migrating less frequently accessed data to HDD to optimize SSD performance and lifespan.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If metadata is stored using conventional DRAM + HDD architecture, then storage capacity is sufficient, but metadata operation performance deteriorates due to high I/O delay of HDD
Solution Approach 1:
The invention segments the storage system into three distinct tiers: DRAM for hot metadata, SSD for warm metadata, and HDD for cold metadata. This segmentation allows each storage medium to be optimized for its specific function, with DRAM providing fast access for frequently accessed metadata, SSD providing balanced performance for moderately accessed metadata, and HDD providing inexpensive bulk storage for rarely accessed metadata, thereby resolving the contradiction between storage capacity and operation performance
Solution Approach 2:
The invention implements dynamic metadata migration between storage tiers based on access frequency and heat classification. Metadata can be moved from HDD to SSD or DRAM when access patterns change, and vice versa when access frequency decreases. This dynamic adjustment allows the system to adapt to changing workloads and maintain optimal performance while efficiently utilizing storage resources across all three tiers
2Productivity
If SSD is used for metadata storage, then metadata operation performance improves, but SSD service life deteriorates due to small granularity random writes
Solution Approach 1:
The invention applies different storage qualities to different metadata based on their access characteristics. Hot metadata that requires frequent access is stored in DRAM with high write endurance, while cold metadata is stored in HDD where write endurance is less critical. For metadata in SSD, the system uses large granularity sequential writes for cold metadata migration, which significantly extends SSD service life compared to small granularity random writes, while still maintaining good operation performance
Solution Approach 2:
The invention introduces DRAM as an intermediary layer between the upper file system and the SSD. The DRAM buffer absorbs and coalesces small granularity random write requests from the file system, converting them into larger sequential write operations when flushing to SSD. This intermediary mechanism protects the SSD from the harmful effects of random writes while maintaining the performance benefits of SSD storage
3Loss of time
If SSD is used for metadata storage, then I/O delay is reduced, but storage cost increases due to high SSD price and small storage capacity
Solution Approach 1:
The invention adds a temporal dimension to the storage hierarchy by classifying metadata into hot, warm, and cold categories based on access time and frequency. This temporal classification enables the system to use expensive high-speed storage (SSD and DRAM) for time-critical hot metadata while using inexpensive bulk storage (HDD) for time-insensitive cold metadata, thereby optimizing the balance between I/O delay and storage cost
4Device complexity
If metadata is stored in directory tree or database structure, then storage organization is simple, but I/O mode becomes small granularity random I/O which is not suitable for SSD
Solution Approach 1:
The invention changes the I/O parameter from small granularity random I/O to large granularity sequential I/O for metadata operations on SSD. By organizing metadata in sequential storage units and using sequential write operations, the system transforms the I/O pattern to match SSD's optimal performance characteristics, thereby significantly improving SSD performance utilization while maintaining relatively simple storage organization through the sequential structure
Data Source
AI summary
The present invention provides a metadata storage system and a management method for a cluster file system. The metadata of the cluster file system is stored by using a DRAM + SSD + HDD three-tier architecture, and the metadata is written sequentially into a metadata buffer in the DRAM based on write-in time. When a metadata buffer fully written, it is written back into the SSD sequentially to form a new metadata storage unit, and corresponding metadata index information in a header structure thereof is stored into the DRAM in a form of a metadata index unit.