Abstract: Herein is an accelerated interface between a database server and a storage area network (SAN). Persistent torage being managed for a database is spread across a number of storage buckets. Global distributed storage metadata is used only for tracking the location of storage buckets on different storage servers. With this approach, a very small amount of memory is needed at a global distributed level to maintain the map. Each storage bucket can have any number of mirrored replicas for further increasing speed and reliability. A database server contains a storage bucket map in memory, and uses the map to do database online transaction processing (OLTP) I/O and smart (i.e. offloaded) database operations on storage. This allows for direct I/O between database server and storage server with lower latency and without using slow and remote middleware such as a logical unit number (LUN) metadata server on a separate network element.

Abstract: Region summaries of database data are stored in persistent memory of a storage cell. Because the region summaries are stored in persistent memory, when a storage cell is powered off and data in volatile memory is not retained, region summaries are nevertheless preserved in persistent memory. When the storage cell comes online, the region summaries already exist and may be used without the delay attendant to regenerating the region summaries stored in volatile memory.

Abstract: Techniques are described for providing one or more clients with direct access to cached data blocks within a persistent memory cache on a storage server. In an embodiment, a storage server maintains a persistent memory cache comprising a plurality of cache lines, each of which represent an allocation unit of block-based storage. The storage server maintains an RDMA table that include a plurality of table entries, each of which maps a respective client to one or more cache lines and a remote access key. An RDMA access request to access a particular cache line is received from a storage server client. The storage server identifies access credentials for the client and determines whether the client has permission to perform the RDMA access on the particular cache line. Upon determining that the client has permissions, the cache line is accessed from the persistent memory cache and sent to the storage server client.

Abstract: Described is an improved approach to implement fast detection of node death. Instead of just relying on multiple heart beats to fail in order to determine whether a node is dead, the present approach performs an on demand validation using RDMA to determine whether the node is reachable, where the approach of using RDMA is significantly faster than the heartbeat approach.

Abstract: Techniques related to a server-side extension of client-side caches are provided. A storage server computer receives, from a database server computer, an eviction notification indicating that a data block has been evicted from the database server computer's cache. The storage server computer comprises a memory hierarchy including a volatile cache and a persistent cache. Upon receiving the eviction notification, the storage server computer retrieves the data block from the persistent cache and stores it in the volatile cache. When the storage server computer receives, from the database server computer, a request for the data block, the storage server computer retrieves the data block from the volatile cache. Furthermore, the storage server computer sends the data block to the database server computer, thereby causing the data block to be stored in the database server computer's cache. Still further, the storage server computer evicts the data block from the volatile cache.

Abstract: A method and an apparatus for implementing a buffer cache for a persistent file system in a non-volatile memory is provided. A set of data is maintained in one or more extents in a non-volatile random-access memory (NVRAM) of a computing device. At least one buffer header is allocated in a dynamic random-access memory (DRAM) of the computing device. In response to a read request by a first process executing on the computing device to access one or more first data blocks in a first extent of the one or more extents, the first process is granted direct read access of the first extent in the NVRAM. A reference to the first extent in the NVRAM is stored in a first buffer header. The first buffer header is associated with the first process. The first process uses the first buffer header to directly access the one or more first data blocks in the NVRAM.

Abstract: Region summaries of database data are stored in persistent memory of a storage cell. Because the region summaries are stored in persistent memory, when a storage cell is powered off and data in volatile memory is not retained, region summaries are nevertheless preserved in persistent memory. When the storage cell comes online, the region summaries already exist and may be used without the delay attendant to regenerating the region summaries stored in volatile memory.

Abstract: Techniques are provided to adjust the behavior of a cache based on a count of cache misses for items recently evicted. In an embodiment, a computer responds to evicting a particular item (PI) from a cache by storing a metadata entry for the PI into memory. In response to a cache miss for the PI, the computer detects whether or not the metadata entry for the PI resides in memory. When the metadata entry for the PI is detected in memory, the computer increments a victim hit counter (VHC) that may be used to calculate how much avoidable thrashing is the cache experiencing, which is how much thrashing would be reduced if the cache were expanded. Either immediately or arbitrarily later, the computer adjusts a policy of the cache based on the VHC's value. For example, the computer may adjust the capacity of the cache based on the VHC.

Abstract: Techniques are described for providing one or more clients with direct access to cached data blocks within a persistent memory cache on a storage server. In an embodiment, a storage server maintains a persistent memory cache comprising a plurality of cache lines, each of which represent an allocation unit of block-based storage. The storage server maintains an RDMA table that include a plurality of table entries, each of which maps a respective client to one or more cache lines and a remote access key. An RDMA access request to access a particular cache line is received from a storage server client. The storage server identifies access credentials for the client and determines whether the client has permission to perform the RDMA access on the particular cache line. Upon determining that the client has permissions, the cache line is accessed from the persistent memory cache and sent to the storage server client.

Abstract: The techniques described herein limit client utilization of a parallel-access storage device. Specifically, client utilization of a particular storage device is estimated using I/O cost metrics to estimate the costs of I/O requests from the client to the particular storage device. The I/O cost metrics are determined based on calibration-based system performance data, which represents a system-wide measure of storage device performance for a system in which the particular storage device resides. The calibration-based system performance data includes one or both of composite throughput data and composite IOPS data for multiple parallel-access devices in the system. The cost estimates for I/O requests issued from a client to a parallel-access device are tracked in a total cost estimate for the client. Client utilization of the storage device, as tracked by the total cost estimate for the client, is limited to a percentage of the total estimated bandwidth of the storage device.

Abstract: Region summaries of database data are stored in persistent memory of a storage cell. Because the region summaries are stored in persistent memory, when a storage cell is powered off and data in volatile memory is not retained, region summaries are nevertheless preserved in persistent memory. When the storage cell comes online, the region summaries already exist and may be used without the delay attendant to regenerating the region summaries stored in volatile memory.

Abstract: Data blocks are cached in a persistent cache (“NV cache”) allocated from as non-volatile RAM (“NVRAM”). The data blocks may be accessed in place in the NV cache of a “source” computing element by another “remote” computing element over a network using remote direct memory access (“RMDA”). In order for a remote computing element to access the data block in NV cache on a source computing element, the remote computing element needs the memory address of the data block within the NV cache. For this purpose, a hash table is stored and maintained in RAM on the source computing element. The hash table identifies the data blocks in the NV cache and specifies a location of the cached data block within the NV cache.

Abstract: Described is an improved approach to implement fast detection of node death. Instead of just relying on multiple heart beats to fail in order to determine whether a node is dead, the present approach performs an on demand validation using RDMA to determine whether the node is reachable, where the approach of using RDMA is significantly faster than the heartbeat approach.

Abstract: A storage system communicatively coupled to a database management system (DBMS performs storage-side scanning of data sources that are not stored in native database storage format of the DBMS. Data sources for external tables are accessible in a storage system referred to as a distributed data access system (DDAS), e.g. a Hadoop Distributed File System. To execute a query that references an external table, a DBMS first generates an execution plan. The DDAS supplies the DBMS with information that specifies each portion of the data source, and specifies which data node to use to access the portion. The DBMS sends a request for each portion to the respective data node, requesting that the data node generate rows from data in the portion. The request may specify scanning criteria, specifying one or more columns to project and/or filter on, and code modules for the data node to execute to generate records.

Abstract: Techniques related to failover to the secondary storage server from a primary storage server of a database server without degrading the performance of servicing storage requests for client applications are provided. In an embodiment, the secondary storage server receives, from the database server, an eviction notification indicating that a set of data blocks has been evicted from a cache. The secondary storage server's memory hierarchy includes a secondary cache and a secondary persistent storage that stores a second copy of the set of data blocks. The secondary storage server persistently stores a copy of data, which is also persistently stored on the primary storage server, which includes a first copy of the set of data blocks.

Abstract: A shared storage architecture persistently stores database files in non-volatile random access memories (NVRAMs) of computing nodes of a multi-node DBMS. The computing nodes of the multi-node DBMS not only collectively store database data on NVRAMs of the computing nodes, but also host database server instances that process queries in parallel, host database sessions and database processes, and together manage access to a database stored on the NVRAMs of the computing nodes. To perform a data block read operation from persistent storage, a data block may be transferred directly over a network between NVRAM of a computing node that persistently stores the data block to a database buffer in non-volatile RAM of another computing node that requests the data block. The transfer is accomplished using remote direct memory access (“RDMA).

Abstract: A method and apparatus for implementing a buffer cache for a persistent file system in non-volatile memory is provided. A set of data is maintained in one or more extents in non-volatile random-access memory (NVRAM) of a computing device. At least one buffer header is allocated in dynamic random-access memory (DRAM) of the computing device. In response to a read request by a first process executing on the computing device to access one or more first data blocks in a first extent of the one or more extents, the first process is granted direct read access of the first extent in NVRAM. A reference to the first extent in NVRAM is stored in a first buffer header. The first buffer header is associated with the first process. The first process uses the first buffer header to directly access the one or more first data blocks in NVRAM.

Abstract: The techniques described herein limit client utilization of a parallel-access storage device. Specifically, client utilization of a particular storage device is estimated using I/O cost metrics to estimate the costs of I/O requests from the client to the particular storage device. The I/O cost metrics are determined based on calibration-based system performance data, which represents a system-wide measure of storage device performance for a system in which the particular storage device resides. The calibration-based system performance data includes one or both of composite throughput data and composite IOPS data for multiple parallel-access devices in the system. The cost estimates for I/O requests issued from a client to a parallel-access device are tracked in a total cost estimate for the client. Client utilization of the storage device, as tracked by the total cost estimate for the client, is limited to a percentage of the total estimated bandwidth of the storage device.

Abstract: A method and apparatus for intelligent network resource manager for distributed computing systems is provided. A first priority is assigned to a first virtual channel set that includes at least two virtual channels of a plurality of virtual channels associated with a physical communication channel. A second priority is assigned to a second virtual channel set that includes at least one virtual channel of the plurality of virtual channels. The first virtual channel set has more virtual channels than the second virtual channel set. Outbound messages of the first priority are directed to virtual channels of the first virtual channel set. Outbound messages of the second priority are directed to virtual channels of the second virtual channel set. The virtual channels are processed in a round-robin order, where processing includes sending the outbound messages over the physical communication channel.

Abstract: A shared storage architecture persistently stores database files in non-volatile random access memories (NVRAMs) of computing nodes of a multi-node DBMS. The computing nodes of the multi-node DBMS not only collectively store database data on NVRAMs of the computing nodes, but also host database server instances that process queries in parallel, host database sessions and database processes, and together manage access to a database stored on the NVRAMs of the computing nodes. To perform a data block read operation from persistent storage, a data block may be transferred directly over a network between NVRAM of a computing node that persistently stores the data block to a database buffer in non-volatile RAM of another computing node that requests the data block. The transfer is accomplished using remote direct memory access (“RDMA).