Abstract: Techniques and mechanisms are provided to allow for selective optimization, including deduplication and/or compression, of portions of files and data blocks. Data access is monitored to generate a heat index for identifying sections of files and volumes that are frequently and infrequently accessed. These frequently used portions may be left non-optimized to reduce or eliminate optimization I/O overhead. Infrequently accessed portions can be more aggressively optimized.

Abstract: A method, computer readable, and apparatus for read-ahead prediction of subsequent requests to send data between a client coupled to a server via a network includes receiving at a traffic management device a request for a part of at least one of a data file and metadata. The traffic management device selects from two or more of a sequential prediction engine, an expert prediction engine and a learning prediction engine to predict a read-ahead of the at least one of the data file and metadata. One or more additional read-ahead parts of the at least one of the data file and metadata are determined with the traffic management device based on the selecting.

Abstract: A system provides file aware block level deduplication in a system having multiple clients connected to a storage subsystem over a network such as an Internet Protocol (IP) network. The system includes client components and storage subsystem components. Client components include a walker that traverses the namespace looking for files that meet the criteria for optimization, a file system daemon that rehydrates the files, and a filter driver that watches all operations going to the file system. Storage subsystem components include an optimizer resident on the nodes of the storage subsystem. The optimizer can use idle processor cycles to perform optimization. Sub-file compression can be performed at the storage subsystem.

Abstract: Mechanisms are provided for efficiently improving a dictionary used for data deduplication. Dictionaries are used to hold hash key and location pairs for deduplicated data. Strong hash keys prevent collisions but weak hash keys are more computation and storage efficient. Mechanisms are provided to use both a weak hash key and a strong hash key. Weak hash keys and corresponding location pairs are stored in an improved dictionary while strong hash keys are maintained with the deduplicated data itself. The need for having uniqueness from a strong hash function is balanced with the deduplication dictionary space savings from a weak hash function.

Abstract: Mechanisms are provided for performing efficient compression and deduplication of data segments. Compression algorithms are learning algorithms that perform better when data segments are large. Deduplication algorithms, however, perform better when data segments are small, as more duplicate small segments are likely to exist. As an optimizer is processing and storing data segments, the optimizer applies the same compression context to compress multiple individual deduplicated data segments as though they are one segment. By compressing deduplicated data segments together within the same context, data reduction can be improved for both deduplication and compression. Mechanisms are applied to compensate for possible performance degradation.

Abstract: Mechanisms are provided for maintaining variable size deduplication segments using fixed size extents. Variable size segments are identified and maintained in a datastore suitcase. Duplicate segments need not be maintained redundantly but can be managed by updating reference counts associated with the segments in the datastore suitcase. Segments are maintained using fixed size extents. A minor increase in storage overhead removes the need for inefficient recompaction when a segment is removed from the datastore suitcase. Fixed size extents can be reallocated for storage of new segments.

Abstract: Mechanisms are provided for accelerated data deduplication. A data stream is received an input interface and maintained in memory. Chunk boundaries are detected and chunk fingerprints are calculated using a deduplication accelerator while a processor maintains a state machine. A deduplication dictionary is accessed using a chunk fingerprint to determine if the associated data chunk has previously been written to persistent memory. If the data chunk has previously been written, reference counts may be updated but the data chunk need not be stored again. Otherwise, datastore suitcases, filemaps, and the deduplication dictionary may be updated to reflect storage of the data chunk. Direct memory access (DMA) addresses are provided to directly transfer a chunk to an output interface as needed.

Abstract: A system provides file aware block level deduplication in a system having multiple clients connected to a storage subsystem over a network such as an Internet Protocol (IP) network. The system includes client components and storage subsystem components. Client components include a walker that traverses the namespace looking for files that meet the criteria for optimization, a file system daemon that rehydrates the files, and a filter driver that watches all operations going to the file system. Storage subsystem components include an optimizer resident on the nodes of the storage subsystem. The optimizer can use idle processor cycles to perform optimization. Sub-file compression can be performed at the storage subsystem.

Abstract: Mechanisms are provided for performing efficient compression and deduplication of data segments. Compression algorithms are learning algorithms that perform better when data segments are large. Deduplication algorithms, however, perform better when data segments are small, as more duplicate small segments are likely to exist. As an optimizer is processing and storing data segments, the optimizer applies the same compression context to compress multiple individual deduplicated data segments as though they are one segment. By compressing deduplicated data segments together within the same context, data reduction can be improved for both deduplication and compression. Mechanisms are applied to compensate for possible performance degradation.

Abstract: Techniques and mechanisms are provided to instantly clone active files including active optimized files. When a new instance of an active file is created, a new stub is generated in the user namespace and a block map file is cloned. The block map file includes the same offsets and location pointers that existed in the original block map file. No user file data needs to be copied. If the cloned file is later modified, the behavior can be same as what happens when a de-duplicated file is modified.

Abstract: Mechanisms are provided for efficiently detecting segments for deduplication. Data is analyzed to determine file types and file components. File types such as images may have optimal data segment boundaries set at the file boundaries. Other file types such as container files are delayered to extract objects to set optimal data segment boundaries based on file type or based on the boundaries of the individual objects. Storage of unnecessary information is minimized in a deduplication dictionary while allowing for effective deduplication.

Abstract: Mechanisms are provided for accelerated data deduplication. A data stream is received an input interface and maintained in memory. Chunk boundaries are detected and chunk fingerprints are calculated using a deduplication accelerator while a processor maintains a state machine. A deduplication dictionary is accessed using a chunk fingerprint to determine if the associated data chunk has previously been written to persistent memory. If the data chunk has previously been written, reference counts may be updated but the data chunk need not be stored again. Otherwise, datastore suitcases, filemaps, and the deduplication dictionary may be updated to reflect storage of the data chunk. Direct memory access (DMA) addresses are provided to directly transfer a chunk to an output interface as needed.

Abstract: Mechanisms are provided for data transfer reduction in scale out architectures. When a compute node receives a write input/output (I/O) request for a data stream, the compute node separates the data stream into chunks and generates fingerprints for the individual chunks. Fingerprints are then sent to a scale out node and compared to fingerprints of chunks already maintained at the scale out node. Write data transfers are only made for chunks not already maintained at the scale out node. For a read I/O request for a data stream, fingerprints for chunks of the data stream are requested by the compute node from a scale out node. Fingerprints received are compared to fingerprints of chunks already maintained at the compute node and read data transfers are only made for chunks not already maintained at the compute node.

Abstract: Deduplication dictionaries are used to maintain data chunk identifier and location pairings in a deduplication system. When access to a particular data chunk is requested, a deduplication dictionary is accessed to determine the location of the data chunk and a datastore is accessed to retrieve the data chunk. However, deduplication dictionaries are large and typically maintained on disk, so dictionary access is expensive. Techniques and mechanisms of the present invention allow prefetches or read aheads of datastore (DS) headers. For example, if a dictionary hit results in datastore DS(X), then headers for DS (X+1), DS (X+2), DS(X+read-ahead-window) are prefetched ahead of time. These datastore headers are cached in memory, and indexed by datastore identifier. Before going to the dictionary, a lookup is first performed in the cached headers to reduce deduplication data access request latency.

Abstract: Mechanisms are provided for efficiently improving a dictionary used for data deduplication. Dictionaries are used to hold hash key and location pairs for deduplicated data. Strong hash keys prevent collisions but weak hash keys are more computation and storage efficient. Mechanisms are provided to use both a weak hash key and a strong hash key. Weak hash keys and corresponding location pairs are stored in an improved dictionary while strong hash keys are maintained with the deduplicated data itself. The need for having uniqueness from a strong hash function is balanced with the deduplication dictionary space savings from a weak hash function.

Abstract: Mechanisms are provided for efficiently improving a dictionary used for data deduplication. Dictionaries are used to hold hash key and location pairs for deduplicated data. Strong hash keys prevent collisions but weak hash keys are more computation and storage efficient. Mechanisms are provided to use both a weak hash key and a strong hash key. Weak hash keys and corresponding location pairs are stored in an improved dictionary while strong hash keys are maintained with the deduplicated data itself. The need for having uniqueness from a strong hash function is balanced with the deduplication dictionary space savings from a weak hash function.

Abstract: Techniques and mechanisms provide a storage optimization manager. Data may be optimized and maintained on various nodes in a cluster. Particular nodes may be overburdened while other nodes remain relatively unused. Techniques are provided to efficiently optimize data onto nodes to enhance operational efficiency. Data access requests for optimized data are monitored and managed to allow for intelligent maintenance of optimized data.

Abstract: Techniques and mechanisms are provided for migrating data blocks around a cluster during node addition and node deletion. Migration requires no downtime, as a newly added node is immediately operational while the data blocks are being moved. Blockmap files and deduplication dictionaries need not be updated.

Abstract: Techniques and mechanisms are provided to instantly clone active files including active optimized files. When a new instance of an active file is created, a new stub is generated in the user namespace and a block map file is cloned. The block map file includes the same offsets and location pointers that existed in the original block map file. No user file data needs to be copied. If the cloned file is later modified, the behavior can be same as what happens when a de-duplicated file is modified.

Abstract: Techniques and mechanisms are provided to support live file optimization. Active I/O access to an optimization target is monitored during optimization. Active files need not be taken offline or made unavailable to an application during optimization and retain the ability to support file operations such as read, write, unlink, and truncate while an optimization engine performs deduplication and/or compression on active file ranges.