Netapp, Inc.

Wei Sun

Anil Paul Thoppil

Anne Maria Vasu

Donald Roman Humlicek

Manan Dahyabhai Patel

Santhosh Selvaraj

Theresa Lynn Segura

Meera Odugoudar

Szu-Wen Kuo

Charles D Binford

Alyssa Proulx

Mahmoud K Jibbe

Mark David Olson

Keith Moyer

Martin Parrish

Satish Sangapu

Anne Marie Vasu

Systems and methods are described for a streamlined garbage collection process during which data integrity checking is also performed for a distributed key-value (KV) store utilized by a distributed storage system. According to one embodiment, by making use of full or truncated block IDs (rather than an intermediate probabilistic data structure, such as a Bloom filter) for garbage collection, data integrity checking can be performed concurrently almost for free. During garbage collection, a block ID compare list may be compared to block IDs within the distributed KV store. If a particular block ID is present in the distributed KV store but is missing from the block ID compare list, the corresponding data block represents garbage to be collected. If the particular block ID is present in the block ID compare list but missing from the distributed KV store, a data integrity error has been identified.

Combined garbage collection and data integrity checking for a distributed key-value store

Systems and methods that make use of cluster-level redundancy within a distributed storage system to address various node-level error scenarios are provided. Rather than making use of a generalized one-size-fits-all approach in an effort to reduce complexity, an approach tailored to the node-level error scenario at issue may be performed to avoid doing more than necessary. According to one embodiment, responsive to identification of a failed RAID stripe by a node of a cluster of a distributed storage management system, for each block ID of multiple block IDs associated with the failed RAID stripe, a data block is restored corresponding to the block ID by reading the data block from another node of the cluster having a redundant copy of the data block; and writing the redundant copy of the data block to a storage area of the node that is unaffected by the failed RAID stripe.

Use of cluster-level redundancy within a cluster of a distributed storage management system to address node-level errors

Techniques are provided for implementing garbage collection and bin synchronization for a distributed storage architecture of worker nodes managing distributed storage composed of bins of blocks. As the distributed storage architecture scales out to accommodate more storage and worker nodes, garbage collection used to free unused blocks becomes unmanageable and slow. Accordingly garbage collection is improved by utilizing heuristics to dynamically speed up or down garbage collection and set sizes for subsets of a bin to process instead of the entire bin. This ensures that garbage collection does not use stale information about what blocks are in-use, and ensures garbage collection does not unduly impact client I/O processing or conversely falls behind on garbage collection. Garbage collection can be incorporated into a bin sync process to improve the efficiency of the bin sync process so that unused blocks are not needlessly copied by the bin sync process.

Garbage collection and bin synchronization for distributed storage architecture

Techniques are provided for implementing a hash building process and an append hash building process. The hash building process builds in-memory hash entries for bins of keys stored within sorted logs of a log structured merge tree used to store keys of a key-value store. The in-memory hash entries can be used to identify the starting locations of bins of keys within the log structured merge tree so that a key within a bin can be searched for from the starting location of the bin as opposed to having to search the entire log structured merge tree. The append hash building process builds two hashes that can be used to more efficiently locate keys and/or ranges of keys within an unsorted append log that would otherwise require a time consuming binary search of the entire append log.

In-memory hash entries and hashes used to improve key search operations for keys of a key value store

Systems and methods that make use of cluster-level redundancy within a distributed storage management system to address various node-level error scenarios are provided. Rather than using a generalized one-size-fits-all approach to reduce complexity, an approach tailored to the node-level error scenario at issue may be performed to avoid doing more than necessary. According to one embodiment, after identifying a missing branch of a tree implemented by a KV store of a first node of a cluster of a distributed storage management system, a branch resynchronization process may be performed, including, for each block ID in the range of block IDs of the missing branch (i) reading a data block corresponding to the block ID from a second node of the cluster that maintains redundant information relating to the block ID; and (ii) restoring the block ID within the KV store by writing the data block to the first node.

Techniques are provided for implementing a garbage collection process and a prediction read ahead mechanism to prefetch keys into memory to improve the efficiency and speed of the garbage collection process. A log structured merge tree is used to store keys of key-value pairs within a key-value store. If a key is no longer referenced by any worker nodes of a distributed storage architecture, then the key can be freed to store other data. Accordingly, garbage collection is performed to identify and free unused keys. The speed and efficiency of garbage collection is improved by dynamically adjusting the amount and rate at which keys are prefetched from disk and cached into faster memory for processing by the garbage collection process.

Prefetching keys for garbage collection

Systems and methods that make use of cluster-level redundancy within a distributed storage management system to address various node-level error scenarios are provided. According to one embodiment, a first node of multiple nodes of distributed storage system represented in a form of a cluster of the multiple of nodes, identifies the potential existence of an error associated with a Redundant Array of Independent Disks (RAID) stripe. A list of block identifiers (IDs) associated with the RAID stripe may then be identified. Rather than performing a traditional RAID recovery/reconstruction approach that is resource intensive in nature and that requires an excessive amount of rebuild time, a more efficient RAID stripe resynchronization process may be performed to restore data associated with the RAID stripe.

Systems and methods are described for a streamlined garbage collection process during which data integrity checking is also performed for a distributed key-value (KV) store utilized by a distributed storage management system. According to one embodiment, by making use of full or truncated block IDs (rather than an intermediate probabilistic data structure, such as a Bloom filter) for garbage collection, data integrity checking can be performed concurrently almost for free. During garbage collection, a block ID compare list is compared to block IDs within the distributed KV store. If a particular block ID is present in the distributed KV store but is missing from the block ID compare list, the corresponding data block represents garbage to be collected. If the particular block ID is present in the block ID compare list but missing from the distributed KV store, a data integrity error has been identified.

Systems and methods that make use of cluster-level redundancy within a distributed storage management system to address various node-level error scenarios are provided. Rather than making use of a generalized one-size-fits-all approach in an effort to reduce complexity, an approach tailored to the node-level error scenario at issue may be performed to avoid doing more than necessary. According to one embodiment, responsive to identification of a failed RAID stripe by a node of a cluster of a distributed storage management system, for each block ID of multiple block IDs associated with the failed RAID stripe, a data block is restored corresponding to the block ID by reading the data block from another node of the cluster having a redundant copy of the data block; and writing the redundant copy of the data block to a storage area of the node that is unaffected by the failed RAID stripe.

Techniques are provided for implementing a defragmentation process during a merge operation performed by a re-compaction process upon a log structured merge tree. The log structured merge tree is used to store keys of key-value pairs within a key-value store. As the log structured merge tree fills with keys over time, the re-compaction process is performed to merge keys down to lower levels of the log structured merge tree to re-compact the keys. Re-compaction can result in fragmentation because there is a lack of spatial locality of where the re-compaction operations re-writes the keys within storage. Fragmentation increases read and write amplification when accessing the keys stored in different locations within the storage. Accordingly, the defragmentation process is performed during a last merge operation of the re-compaction process in order to store keys together within the storage, thus reducing read and write amplification when accessing the keys.

Defragmentation for log structured merge tree to improve read and write amplification

Systems and methods that make use of cluster-level redundancy within a distributed storage management system to address various node-level error scenarios are provided. Rather than using a generalized one-size-fits-all approach to reduce complexity, an approach tailored to the node-level error scenario at issue may be performed to avoid doing more than necessary. According to one embodiment, responsive to identifying a missing branch of a tree implemented by a KV store of a first node of a cluster of a distributed storage management system, a branch resynchronization process may be performed, including, for each block ID in the range of block IDs of the missing branch (i) reading a data block corresponding to the block ID from a second node of the cluster that maintains redundant information relating to the block ID; and (ii) restoring the block ID within the KV store by writing the data block to the first node.

Systems and methods that make use of cluster-level redundancy within a distributed storage management system to address various node-level error scenarios are provided. According to one embodiment, a KV store of a node of a cluster of a distributed storage management system manages storage of data blocks as values and corresponding block IDs as keys. Data integrity errors are reported to the first node in the form of a list of missing block IDs that are in use but missing from the KV store. A metadata resynchronization process may then be caused to be performed, including for each block ID in the list of missing block IDs: (i) reading a data block corresponding to the block ID from another node of the cluster that maintains redundant information relating to the block ID; and (ii) restoring the block ID within the KV store by writing the data block to the node.

A system, method, and machine-readable storage medium for performing garbage collection in a distributed storage system are provided. In some embodiments, an efficiency level of a garbage collection process is monitored. The garbage collection process may include removal of one or more data blocks of a set of data blocks that is referenced by a set of content identifiers. The set of slice services and the set of data blocks may reside in a cluster, and a set of probabilistic filters (e.g., Bloom filters) may indicate whether the set of data blocks is in-use. At least one parameter of a probabilistic filter of the set of probabilistic filters may be adjusted (e.g., increased or reduced) if the efficiency level is below the efficiency threshold. Garbage collection may be performed on the set of data blocks in accordance with the set of probabilistic filters.

Growing community of inventors

Boulder, CO, United States of America

Wei Sun