Panasas, Inc.

Mark C Holland

James D Zelenka

Brent B Welch

Garth Gibson

Denis M Serenyi

A method for allowing client computers in a distributed object-based data storage system to interpret higher revision data structures. In one embodiment, clients request layout maps of objects to be accessed from respective storage managers in the system and then construct and execute directed acyclic graphs (DAG) to accomplish desired I/O operations for data access. When a client computer running an older software revision is not able to interpret a particular data organization supplied by the storage manager that was designed after that client's software was installed, the client sends to the storage manager a description of the desired data access operation. The storage manager, in turn, interprets a data layout scheme for a client on a per-access basis and returns to that client a DAG containing a data structure that describes the sequence of I/O operations necessary to effect the data access. The client can then execute the storage manager-supplied DAG to carry out the data access.

Allowing client systems to interpret higher-revision data structures in storage systems

A methodology wherein two different types of locks are used by a storage manager when multiple clients wish to access a particular redundantly-stored file. Simple byte-range based mutual exclusion (or mutex) locks are granted by the storage manager for data writes/updates to the file when the file is in the fault-free state, and individual readers/writers (R/W) locks are granted by the storage manager when the file is in the degraded state. No read locks are required of clients when the file object is in the fault-free state. During the fault-free state of the file object, when exactly one client is writing to the file object, the storage manger grants that file object a whole-file lock valid over the entire file object. Each client may have a client lock manager that interacts with appropriate storage manager lock manager to request and obtain necessary locks. These various locking mechanisms reduce lock-related network traffic in a data storage system.

Using whole-file and dual-mode locks to reduce locking traffic in data storage systems

A hybrid quorum/consensus and primary-backup fault-tolerance model in an object-based distributed data storage system. When a primary manager fails, a hierarchy of network entities is established in which a group of realm managers first authorizes a failure-handling event through quorum/consensus and a backup manager (for the failing primary manager) then executes the decision of the quorum of realm managers. The realm managers, operating by consensus, determine whether (a) the primary manager can indeed be asserted to be down, and (b) whether there is a quorum of realm managers in agreement on this decision. If both are true, a master realm manager instructs the backup manager to proceed to execute the necessary steps to become the primary manager and function as the primary until the original primary manager is brought back into service. The hybrid fault-tolerance approach handles both single unit failures and network partitions in a unified way, without creating a single cluster out of the fault domain.

Hybrid quorum/primary-backup fault-tolerance model

A data storage methodology wherein fault tolerance is accomplished by generation of multiple canonical names for each primary manager active in the data storage system and then balancing the primary manager's failure-induced workload across the multiple backup managers using the canonical names. A qualified name for each backup manager is derived from a canonical name assigned to the primary manager. During fault-free state of operation, all the qualified names are assigned to the primary manager, which serves all its qualified names. When that primary manager fails, the work it was supporting is distributed among its backup managers using the qualified names for those backups. The distribution of primary manager's workload among multiple backups eliminates the severe performance bottlenecks present in the traditional approach of keeping one backup manager for a primary manager and keeping that backup manager idle during the fault-free state of its primary manager.

Distributing manager failure-induced workload through the use of a manager-naming scheme

A system and method to perform a system-wide file system image without time smear in a distributed object-based data storage system. A realm manager is elected as an image master using the Distributed Consensus Algorithm to execute image-taking. All pending write capabilities are invalidated prior to taking the system-wide file system image so as to quiesce the realm and prepare the storage system for the system-wide image. Once the system is quiesced, the image master instructs each storage manager in the system to clone each live object group contained therein without explicitly cloning any objects contained in such object group. In one embodiment, a file manager copies an object in the system before a write operation is performed on that object after the image is taken. Neither the cloning operation nor the copying operation update any directory objects in the system. At run time, a client application may use a mapping scheme to access objects contained in the system-wide image.

Internally consistent file system image in distributed object-based data storage

A data storage methodology wherein a data file is initially stored in a format consistent with RAID-1 and RAID-X and then migrated to a format consistent with RAID-X and inconsistent with RAID-1 when the data file grows in size beyond a certain threshold. Here, RAID-X refers to any non-mirrored storage scheme employing XOR-based error correction coding (e.g., a RAID-5 configuration). Each component object (including the data objects and the parity object) for the data file is configured to be stored in a different stripe unit per object-based secure disk. Each stripe unit may store, for example, 64 KB of data. So long as the data file does not grow beyond the size threshold of a stripe unit (e.g., 64 KB), the parity stripe unit contains a mirrored copy of the data stored in one of the data stripe units because of the exclusive-ORing of the input data with 'all zeros' assumed to be contained in empty or partially-filled stripe units. When the file grows beyond the size threshold, the parity stripe unit starts storing parity information instead of a mirrored copy of the file data. Thus, the data file can be automatically migrated from a format consistent with RAID-1 and RAID-X to a format consistent with RAID-X and inconsistent with RAID-1 without the necessity to duplicate or rewrite the stored data.

Growing community of inventors

Pittsburgh, PA, United States of America

Mark C Holland