|
Object-Based Architecture
|
|
Extreme scalability
|
Lustre's object-based storage architecture scales to tens of thousands of nodes, petabytes of data, and billions of files. The highly scalable distributed file system combines open standards, Linux, an open API, and innovative protocols to create the world’s largest network-neutral data storage and retrieval system.
|
|
|
Shatters performance limitations
|
Applying intelligence throughout its architecture, Lustre turns commodity hardware into smart storage devices that manage data objects. The objects are dynamically distributed horizontally across the servers, shattering the performance limitations of traditional storage systems.
|
|
|
Unique approach
|
Building a Lustre cluster requires a Lustre MetaData Server (MDS) and Lustre Object Storage Servers (OSSs), each with disk storage. A pool of client systems access these servers through one of many supported networks. Lustre file operations bypass the MetaData server completely and full utilize the parallel data paths to all OSSs in the cluster. This unique approach—separating metadata operations—results in significantly enhanced performance.
Like other Unix and Linux file systems, Lustre files are represented by inodes. But a key difference in Lustre is that its inodes simply contain references to the objects storing the file data.
|
|
Production-Quality Reliability
|
|
Meeting uptime requirements
|
Lustre is currently deployed in both large and small clusters around the world, meeting the uptime requirements of business-critical and national-security applications. In fact, more than 100 teraflops worldwide are deployed in production environments where Lustre is used to manage data.
|
|
|
No point of failure
|
The system delivers production-quality stability and failover, with no single point of failure. Lustre organizes all servers in active-active failover pairs. Together with protocol interoperability between versions, live cluster upgrades are now routine.
|
|
|
High-availability configuration
|
Lustre's just-mount configuration feature simplifies configuration by routinely formatting and mounting server devices and then aggregating them into a global high-availability cluster file system.
|
|
Proven Performance
|
|
Supercomputer power
|
Lustre is designed for data-intensive applications, and delivers dramatically increased throughput and I/O through intelligent serialization and separation of metadata operations from data manipulation. The Lustre system powers most of the world’s largest Linux supercomputers and is the first production-tested, object-based Linux cluster file system.
|
|
|
Innovative protocols
|
Lustre employs a distributed lock manager to handle access to files and directories and synchronize updates, improving on the metadata journaling approach used by most modern file systems.
|
|
|
Intent-based locking
|
To reduce bottlenecks and increase overall data throughput, Lustre uses an intent-based locking mechanism, where file and directory lock requests also provide information about the reason the lock is being requested. For example, if a directory lock is being requested to create a new, unique file, Lustre handles this as a single request. In other file systems, this action requires multiple network requests for lookup, creation, opening and locking.
|
|
|
Extreme parallel computing
|
The Lustre lock manager automatically adapts its policies to minimize overhead for its current application. Files that are being used by a single node are covered by a single lock, eliminating additional lock overhead. Nodes sharing files get the largest possible locks, which still allows all nodes to write at full speed.
|
|
Heterogeneous Networking
|
|
Open source, open standards
|
Lustre was developed and maintained as open source software under the GNU General Public License (GPL). The system features an open networking protocol and POSIX file system semantics, ensuring broad support for industry-standard platforms and heterogeneous networking environments.
|
|
|
Supports a variety of networks
|
Lustre's network architecture provides flexible support for a wide range of networks and high-performance features. Lustre interoperates with network vendor-supplied libraries through Lustre Network Drivers (LNDs) that utilize advanced features such as Remote Direct Memory Access (EDMA) OS-bypass for parallel I/O, and vector I/O for efficient bulk data movement. LNDs exist for many networks, including TCP, Quadrics Elan, many flavors of InfiniBand, and Myrinet GM, each with performance exceptionally close to the raw device throughput.
|