Amazon Technologies, Inc.

Quanta Computer Incorporated

Matteo Frigo

Matti Juhani Oikarinen

Pradeep Vincent

Jacob A Strauss

Wayne William Duso

Michael Robert Frasca

Colin Laird Lazier

Bryan James Donlan

Douglas Stewart Laurence

Lawrence Colm Stewart

Michael Anthony Nuss

Joshua Samuel Hendrickson

Alex Haugland

James Christopher Sorenson

Allan Henry Vermeulen

Daniel Nussbaum

Xiaobin Wu

Iii James Christopher Sorenson

Techniques described and suggested include systems and processes for handling data and client lock management in multi-tenant storage systems, such as distributed storage systems. In some embodiments, data lock and client lock records are stored in connection with a plurality of data stores on which client data resides. In some embodiments, a storage hub operably connected to the data stores determines and retain client lease information. In some embodiments, the client lease information is used in conjunction with the data lock and client lock records to effect and determine the validity of locks on various data elements.

Data locking and state management on distributed storage systems

Erasure encoded fragments are generated by an erasure encoding scheme, represented by an erasure encoding matrix, operating on a data file. A new erasure encoded fragment may be generated from previously-generated erasure encoded fragments without reconstructing the original data file. Available and valid erasure encoded fragments are identified and a set of those fragments is selected. A composite encoding matrix is generated based upon the selected fragments and the fragment specified to be generated. The composite matrix is applied to the selected fragments to produce a plurality of partial sums. The partial sums are then combined to generate the specified fragment. The partial sums may be produced by different devices so as to distribute the computational workload and/or to reduce network traffic. The integrity of a generated fragment may be verified by generating the specified fragment twice, using two different sets of fragments, and then comparing the two results.

Generation and verification of erasure encoded fragments

A computing resource service provider may operate a data storage service configured to provide data storage for one or more customers of the computing resource service provider. The data storage service may store customer data in one or more replicated state machines, where the replicated state machines comprise a plurality of replicated state machine-shards. The replicated state machine-shards may cause the computer system hosting the replicated state machine-shard to transmit a consensus message to other computer system. The consensus message may include utilization information corresponding to the other computer system. The utilization information may be used to calculate a utilization rate for the replicated state machine usable in block allocation operations.

Block allocation based on server utilization

A client request, formatted in accordance with a file system interface, is received at an access subsystem of a distributed multi-tenant storage service. After the request is authenticated at the access subsystem, an atomic metadata operation comprising a group of file system metadata modifications is initiated, including a first metadata modification at a first node of a metadata subsystem of the storage service and a second metadata modification at a second node of the metadata subsystem. A plurality of replicas of at least one data modification corresponding to the request are saved at respective storage nodes of the service.

Scalable file storage service

An example method to decode an error-correcting code includes receiving a tuple including a seed and a coded packet over a communication channel, and based on the seed, reconstructing a set of pseudorandom coefficients used by an encoder to create the coded packet. The method also includes entering the set of pseudorandom coefficients and the coded packet in a decoding matrix and reducing the decoding matrix to a row echelon form. After reducing the decoding matrix to a row echelon form, the method further includes determining if the decoding matrix is decodable up to a number of rows, reducing the decoding matrix up to the number of rows by backward substitution so a part of the first matrix becomes an identity matrix and a corresponding part of the second matrix comprise source packets, and extracting the source packets from the padded packets.

Error-correcting code

A metadata manager of a file system receives a directory move request, indicating a source directory (SD) and a proposed parent directory (PPD). Each directory in the directory tree is assigned a respective move rank (MR) in accordance with a rank-order preserving rule. The metadata manager checks whether the MR of the SD or the PPD is to be changed to avoid violating the rule. If an MR change is required, one or more directory-pair operations are implemented to determine whether the move can be successfully completed. In each directory-pair operation, the MRs of a pair of directories is examined and an attempt is made to modify one of the MRs of the pair in accordance with the rule. If enough MRs can be changed, the move is completed by adding the SD as a child of the PPD.

Concurrent directory move operations using ranking rules

A computing resource service provider may operate a data storage service configured to provide data storage for one or more customers of the computing resource service provider. The data storage service may store customer data in one or more replicated state machines, where the replicated state machines comprise a plurality of replicated state machines-shards. The replicated state machines-shards may cause the computer system hosting the replicated state machine-shard to transmit a consensus message to other computer system. The consensus message may include utilization information corresponding to the other computer system. The utilization information may be used to calculate a utilization rate for the replicated state machine useable in block allocation operations.

In response to receiving a write request directed to a particular logical block of a storage object, a page of free space (sufficient to accommodate the payload of the write request, but smaller in size than the logical block) of a particular extent that has been selected to store contents of the logical block is allocated. The current size of the extent is smaller than the combined sizes of logical blocks that are mapped to the extent. The page is modified in accordance with a payload indicated in the write request. In response to a subsequent write request directed to the particular extent, a determination is made that the particular extent would violate a free space threshold criterion if the payload of the write request were accommodated, and an extent expansion operation is initiated.

Oversubscribed storage extents with on-demand page allocation

A write request directed to a storage object is received at a distributed file storage service. Based on a variable stripe size selection policy, a size of a particular stripe of storage space to be allocated for the storage object is determined, which differs from the size of another stripe allocated earlier for the same storage object. Allocation of storage for the particular stripe at a particular storage device is requested, and if the allocation succeeds, the contents of the storage device are modified in accordance with the write request.

File storage using variable stripe sizes

An I/O request directed to a portion of a storage object managed at a distributed storage service is received. A congestion control parameter value to be used to schedule a storage operation corresponding to the I/O request is determined. The congestion control parameter is based at least in part on an offset within the storage object to which the I/O request is directed. The storage operation is scheduled in accordance with the congestion control parameter at a selected physical storage device to which the portion of the storage object is mapped.

Congestion control in storage systems

A method is provided to transmit an error-correcting code. The method includes receiving a source packet, advancing a sliding encoder window after receiving the source packet, and generating coded packets by, when the padded packet is encoded for a first time, encoding the source packet alone, and, when the source packet is encoded for a number of times after the first time, encoding the source packet with older source packets in the sliding window.

A connection request from a client is received at an access subsystem node of a distributed storage service. The connection request includes an attempt count parameter indicative of a number of times an establishment of a connection on behalf of the client has been attempted. A workload threshold level based on the attempt count parameter is identified for use in an acceptance decision for the connection request. In response to a determination that a local workload metric of the access subsystem node is below the workload threshold level, the connection is accepted.

Connection balancing using attempt counts at distributed storage systems

A node of a distributed storage service receives a read request from a client, directed to a particular physical page storing at least a portion of a file store object. The node transmits, to the client, a particular operation sequence number (OSN), obtained from a replicated state machine, indicative of an order in which the read request was processed at the node relative to other requests directed to the file store object. The node receives a write request from the client directed to the page. The write request includes the particular OSN. The node determines whether the page has been modified after the particular OSN was obtained. If the page has not been modified since the OSN was obtained, a modification indicated in the write request is performed; otherwise, the write request is rejected.

Conditional writes at distributed storage services

A node of a storage service is selected as a coordinator of a distributed transaction involving multiple page-level modifications. The coordinator identifies other nodes as members of a node chain collectively storing physical data pages at which proposed modifications are to be performed, including a decider node responsible for a decision to commit the transaction. The coordinator generates a transaction preparation message comprising a representation of an order of respective commit decisions associated with the proposed modifications, and transmits the message to a selected node of the chain for a sequential propagation along the chain. Each chain node performs a local commit analysis for its changes and stores a record of its intent to commit. If a decision to commit is reached at the decider, the proposed modifications are completed.

Growing community of inventors

Acton, MA, United States of America

Matteo Frigo