Distributed Storage Metadata for Scalable Failure Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional file systems face scalability limitations due to the use of limited metadata servers, leading to potential data loss in case of server failure and inefficiencies in load distribution, especially in large clusters.
Innovation Solution
A distributed file system architecture utilizing GPUs and non-volatile memory (NVM) to distribute metadata across many servers, enabling scalable and resilient storage with independent bucket ownership and parallel operations, allowing for efficient load balancing and failure recovery.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a centralized metadata server is used, then the file system structure is simple and easy to manage, but the system cannot scale and is vulnerable to single point of failure
Solution Approach 1:
The patent divides the centralized metadata server into multiple distributed metadata servers, each managing a portion of the file system metadata. This segmentation eliminates the single point of failure while maintaining manageable complexity through modular architecture. Each metadata server operates independently but coordinates through the distributed file system protocol.
Solution Approach 2:
The patent changes the architectural parameter from centralized to distributed metadata management. This fundamental parameter change transforms the system from a single-point-of-failure model to a resilient distributed model where metadata is replicated across multiple servers, improving reliability without proportionally increasing operational complexity.
2Reliability
If multiple metadata servers are used in a clustered way, then reliability is improved, but the system cannot scale to large clusters
Solution Approach 1:
The patent segments metadata management into independent metadata servers that each handle specific portions of the file system namespace. This segmentation enables linear scaling - adding more servers increases both reliability and capacity proportionally, allowing the system to scale to large clusters without quadratic complexity increases.
Solution Approach 2:
The patent transitions from a two-dimensional cluster management model to a three-dimensional hierarchical model with metadata servers organized in tiers. This dimensional change allows for more efficient load distribution and failure isolation, enabling scalability to large clusters while maintaining manageable complexity through hierarchical organization.
3Ease of operation
If traditional storage systems are used, then ease of operation is maintained, but load distribution is inefficient in large clusters
Solution Approach 1:
The patent implements self-service load distribution where clients autonomously select metadata servers based on current system state and data location. This eliminates the need for centralized load balancing, maintaining ease of operation from the user perspective while dramatically improving load distribution efficiency through client-side intelligence and distributed decision-making.
Solution Approach 2:
The patent incorporates feedback mechanisms where metadata servers report their status and load information to clients, enabling dynamic load balancing. This feedback loop allows the system to automatically redistribute loads based on real-time conditions, improving productivity without requiring manual intervention or complex centralized control.
Data Source
AI summary
A plurality of computing devices are communicatively coupled to each other via a network, and each of the plurality of computing devices is operably coupled to one or more of a plurality of storage devices. A plurality of failure resilient stripes is distributed across the plurality of storage devices such that each of the plurality of failure resilient stripes spans a plurality of the storage devices. A graphics processing unit is operable to access data files from the failure resilient stripes, while bypassing a kernel page cache. Furthermore, these data files may be accessed in parallel by the graphics processing unit.