General Parallel File System

GPFS is a shared-disk, parallel file system that stripes data across multiple storage devices and servers, allowing many nodes to read and write concurrently. It uses a distributed metadata architecture, with metadata servers managing file attributes and directory structures, while data is stored in 256 KiB blocks by default. The system employs a token-based cache coherence protocol to ensure consistency across nodes, and it supports POSIX semantics, making it compatible with existing applications. GPFS also provides high availability through replication and failover, and it can be deployed on commodity hardware or specialized storage systems.

GPFS originated from IBM's research on parallel file systems in the early 1990s, with the first commercial release in 1998. It was initially developed for AIX and later ported to Linux and Windows. In 2014, IBM rebranded GPFS as IBM Spectrum Scale, integrating it with the Spectrum Storage family. Over the years, GPFS has added features such as information lifecycle management, compression, encryption, and support for object storage via the S3 API. It has been adopted in many of the world's largest supercomputers, including the IBM Blue Gene series and the Summit supercomputer.

GPFS is used in a variety of demanding environments, from scientific research and weather forecasting to financial services and media production. It is the underlying file system for many of the TOP500 supercomputers, providing the I/O performance needed for large-scale simulations and data analytics. GPFS supports multiple access protocols, including NFS, SMB, and POSIX, and it can be integrated with Hadoop and Spark for big data workloads. Its ability to scale linearly with the number of nodes makes it a preferred choice for organizations that need to manage massive datasets with high throughput and low latency.

Beyond its mainstream use, GPFS has several niche features and historical quirks. For instance, GPFS was one of the first file systems to implement a fully distributed locking mechanism, which was later adopted by other parallel file systems. It also supports a 'tiebreaker' node to resolve split-brain scenarios in clustered configurations. In the early 2000s, GPFS was used in the Large Hadron Collider's data storage infrastructure, handling petabytes of particle physics data. Additionally, GPFS has a feature called 'AFM' (Active File Management) that enables asynchronous data replication across geographically distributed clusters, which is often overlooked. The file system's ability to run on a wide range of hardware, from small clusters to massive supercomputers, has made it a versatile tool in both academic and commercial settings.

Clustered file system

A file system that is simultaneously mounted on multiple servers, providing a unified namespace and concurrent access.

Metadata

Data about data, such as file names, timestamps, and permissions, used to manage and locate files.

POSIX

A set of standards for operating system interfaces, ensuring compatibility between Unix-like systems.

Token-based cache coherence

A protocol that uses tokens to manage cache consistency across distributed nodes.