Abstract: Techniques are described for encoding join columns that belong to the same domain with a common dictionary. The tables are encoded with dictionary indexes that make the comparison operation of a join query a quick equality check of two integers and there is no need to compute any hashes during execution. Additionally, the techniques described herein minimize the bloom filter creation and evaluation cost as well because the dictionary indexes serve as hash values into the bloom filter. If the bloom filter is as large as the range of dictionary indexes, then the filter is no longer a probabilistic structure and can be used to filter rows in the probe phase with full certainty without any significant overhead.

Abstract: Techniques are provided for storing in in-memory unit (IMU) in a lower-storage tier and copying the IMU to DRAM when needed for query processing. Techniques are also provided for copying IMUs to lower tiers of storage when evicted from the cache of higher tiers of storage. Techniques are provided for implementing functionality of IMUs within a storage system, to enable database servers to push tasks, such as filtering, to the storage system where the storage system may access IMUs within its own memory to perform the tasks. Metadata associated with a set of data may be used to indicate whether an IMU for the data should be created by the database server machine or within the storage system.

Abstract: Techniques are provided to allow more sophisticated operations to be performed remotely by machines that are not fully functional. Operations that can be performed reliably by a machine that has experienced a hardware and/or software error are referred to herein as Remote Direct Memory Operations or “RDMOs”. Unlike RDMAs, which typically involve trivially simple operations such as the retrieval of a single value from the memory of a remote machine, RDMOs may be arbitrarily complex. The techniques described herein can help applications run without interruption when there are software faults or glitches on a remote system with which they interact.

Abstract: Techniques are described herein for distributing distinct portions of a database object across volatile memories of selected nodes of a plurality of nodes in a clustered database system. The techniques involve storing a unit-to-service mapping that associates a unit (a database object or portion thereof) to one or more database services. The one or more database services are mapped to one or more nodes. The nodes to which a service is mapped may include nodes in disjoint database systems, so long as those database systems have access to a replica of the unit. The database object is treated as in-memory enabled by nodes that are associated with the service, and are treated as not in-memory enabled by nodes that are not associated with the service.

Abstract: A method, apparatus, and system for OZIP, a data compression and decompression codec, is provided. OZIP utilizes a fixed size static dictionary, which may be generated from a random sampling of input data to be compressed. Compression by direct token encoding to the static dictionary streamlines the encoding and avoids expensive conditional branching, facilitating hardware implementation and high parallelism. By bounding token definition sizes and static dictionary sizes to hardware architecture constraints such as word size or processor cache size, hardware implementation can be made fast and cost effective. For example, decompression may be accelerated by using SIMD instruction processor extensions. A highly granular block mapping in optional stored metadata allows compressed data to be accessed quickly at random, bypassing the processing overhead of dynamic dictionaries. Thus, OZIP can support low latency random data access for highly random workloads, such as for OLTP systems.

Abstract: A hashing scheme includes a cache-friendly, latchless, non-blocking dynamically resizable hash index with constant-time lookup operations that is also amenable to fast lookups via remote memory access. Specifically, the hashing scheme provides each of the following features: latchless reads, fine grained lightweight locks for writers, non-blocking dynamic resizability, cache-friendly access, constant-time lookup operations, amenable to remote memory access via RDMA protocol through one sided read operations, as well as non-RDMA access.

Abstract: A hashing scheme includes a cache-friendly, latchless, non-blocking dynamically resizable hash index with constant-time lookup operations that is also amenable to fast lookups via remote memory access. Specifically, the hashing scheme provides each of the following features: latchless reads, fine grained lightweight locks for writers, non-blocking dynamic resizability, cache-friendly access, constant-time lookup operations, amenable to remote memory access via RDMA protocol through one sided read operations, as well as non-RDMA access.

Abstract: A method for accelerating queries using dynamically generated columnar data in a flash cache is provided. In an embodiment, a method comprises a storage device receiving a first request for data that is stored in the storage device in a base major format in one or more primary storage devices. The storage device comprises a cache. The base major format is any one of: a row-major format, a column-major format and a hybrid-columnar format. Based on first one or more criteria, it is determined whether to rewrite the data into rewritten data in a rewritten major format. In response to determining to rewrite the data into rewritten data in a rewritten major format, the storage device rewrites at least a portion of the data into particular rewritten data in the rewritten major format. The rewritten data is stored in the cache.

Abstract: A method and apparatus for efficiently processing data in various formats in a single instruction multiple data (“SIMD”) architecture is presented. Specifically, a method to unpack a fixed-width bit values in a bit stream to a fixed width byte stream in a SIMD architecture is presented. A method to unpack variable-length byte packed values in a byte stream in a SIMD architecture is presented. A method to decompress a run length encoded compressed bit-vector in a SIMD architecture is presented. A method to return the offset of each bit set to one in a bit-vector in a SIMD architecture is presented. A method to fetch bits from a bit-vector at specified offsets relative to a base in a SIMD architecture is presented. A method to compare values stored in two SIMD registers is presented.

Abstract: Techniques are described for materializing pre-computed results of expressions. In an embodiment, a set of one or more column units are stored in volatile or non-volatile memory. Each column unit corresponds to a column that belongs to an on-disk table within a database managed by a database server instance and includes data items from the corresponding column. A set of one or more virtual column units, and data that associates the set of one or more column units with the set of one or more virtual column units, are also stored in memory. The set of one or more virtual column units includes a particular virtual column unit storing results that are derived by evaluating an expression on at least one column of the on-disk table.

Abstract: A hashing scheme includes a cache-friendly, latchless, non-blocking dynamically resizable hash index with constant-time lookup operations that is also amenable to fast lookups via remote memory access. Specifically, the hashing scheme provides each of the following features: latchless reads, fine grained lightweight locks for writers, non-blocking dynamic resizability, cache-friendly access, constant-time lookup operations, amenable to remote memory access via RDMA protocol through one sided read operations, as well as non-RDMA access.

Abstract: Techniques are provided for enabling a requesting entity to retrieve data that is managed by a database server instance from the volatile memory of a server machine that is executing the database server instance. The techniques allow the requesting entity to retrieve the data from the volatile memory of the host server machine without involving the database server instance in the retrieval operation. Because the retrieval does not involve the database server instance, the retrieval may succeed even when the database server instance has stalled or become unresponsive. In addition, direct retrieval of data using the techniques described herein will often be faster and more efficient than retrieval of the same information through conventional interaction with the database server instance.

Abstract: Techniques related to efficient data storage and retrieval using a heterogeneous main memory are disclosed. A database includes a set of persistent format (PF) data that is stored on persistent storage in a persistent format. The database is maintained on the persistent storage and is accessible to a database server. The database server converts the set of PF data to sets of mirror format (MF) data and stores the MF data in a hierarchy of random-access memories (RAMs). Each RAM in the hierarchy has an associated latency that is different from a latency associated with any other RAM in the hierarchy. Storing the sets of MF data in the hierarchy of RAMs includes (1) selecting, based on one or more criteria, a respective RAM in the hierarchy to store each set of MF data and (2) storing said each set of MF data in the respective RAM.

Abstract: Techniques are provided for maintaining data persistently in one format, but making that data available to a database server in more than one format. For example, one of the formats in which the data is made available for query processing is based on the on-disk format, while another of the formats in which the data is made available for query processing is independent of the on-disk format. Data that is in the format that is independent of the disk format may be maintained exclusively in volatile memory to reduce the overhead associated with keeping the data in sync with the on-disk format copies of the data.

Abstract: Techniques are provided for determining costs for alternative execution plans for a query, where at least a portion of the data items required by the query are in in-memory compression-units within volatile memory. The techniques involve maintaining in-memory statistics, such as statistics that indicate what fraction of a table is currently present in in-memory compression units, and the cost of decompressing in-memory compression units. Those statistics are used to determine, for example, the cost of a table scan that retrieves some or all of the necessary data items from the in-memory compression-units.

Abstract: A method and apparatus is provided for optimizing queries received by a database system that relies on an intelligent data storage server to manage storage for the database system. Storing compression units in hybrid columnar format, the storage manager evaluates simple predicates and only returns data blocks containing rows that satisfy those predicates. The returned data blocks are not necessarily stored persistently on disk. That is, the storage manager is not limited to returning disc block images. The hybrid columnar format enables optimizations that provide better performance when processing typical database workloads including both fetching rows by identifier and performing table scans.

Abstract: To prioritize repopulation of in-memory compression units (IMCU), a database server compresses, into an IMCU, a plurality of data units from a database table. In response to changes to any of the plurality of data units within the database table, the database server performs the steps of: (a) invalidating corresponding data units in the IMCU; (b) incrementing an invalidity counter of the IMCU that reflects how many data units within the IMCU have been invalidated; (c) receiving a data request that targets one or more of the plurality of data units of the database table; (d) in response to receiving the data request, incrementing an access counter of the IMCU; and (e) determining a priority for repopulating the IMCU based, at least in part, on the invalidity counter and the access counter.

Abstract: Techniques are provided for maintaining data persistently in one format, but making that data available to a database server in more than one format. For example, one of the formats in which the data is made available for query processing is based on the on-disk format, while another of the formats in which the data is made available for query processing is independent of the on-disk format. Data that is in the format that is independent of the disk format may be maintained exclusively in volatile memory to reduce the overhead associated with keeping the data in sync with the on-disk format copies of the data.

Abstract: Techniques are provided for more efficiently using the bandwidth of the I/O path between a CPU and volatile memory during the performance of database operation. Relational data from a relational table is stored in volatile memory as column vectors, where each column vector contains values for a particular column of the table. A binary-comparable format may be used to represent each value within a column vector, regardless of the data type associated with the column. The column vectors may be compressed and/or encoded while in volatile memory, and decompressed/decoded on-the-fly within the CPU. Alternatively, the CPU may be designed to perform operations directly on the compressed and/or encoded column vector data. In addition, techniques are described that enable the CPU to perform vector processing operations on the column vector values.

Abstract: Techniques are described herein for distributing distinct portions of a database object across volatile memories of selected nodes of a plurality of nodes in a clustered database system. The techniques involve storing a unit-to-service mapping that associates a unit (a database object or portion thereof) to one or more database services. The one or more database services are mapped to one or more nodes. The nodes to which a service is mapped may include nodes in disjoint database systems, so long as those database systems have access to a replica of the unit. The database object is treated as in-memory enabled by nodes that are associated with the service, and are treated as not in-memory enabled by nodes that are not associated with the service.