B+ tree management method based on remote direct nonvolatile memory access

Abstract

The invention relates to a B+ tree management method based on remote direct nonvolatile memory access. According to the method, internal nodes of a complete B+ tree are stored on a DRAM, volatile leaf nodes on the DRAM only contain keywords, data corresponding to the keywords is stored in leaf nodes on a nonvolatile memory, and each volatile leaf node is associated with one nonvolatile leaf node; a client acquires a remote lock through RDMA atomic operation and then sends an operation command of the B+ tree to a server; the server reserves remote operation space in the corresponding nonvolatile leaf nodes; and the client uses an RDMA remote writing technology to directly persist the data into the nonvolatile leaf nodes of the server in a log form. Through the method, the client cooperates with the server in operation on the B+ tree of the nonvolatile memory, so that processing burden on the server is lowered under the condition of concurrent access; and besides, operation is performed on the nonvolatile leaf nodes in the log form, so that writing wear of the NVM is relieved.

Description

technical field [0001] The invention relates to a B+ tree management mechanism, in particular to a B+ tree management method based on remote direct non-volatile memory access. Background technique [0002] As an important part of the storage system function, the index can speed up the search of data. As an index structure of the widely used B-tree class, it is originally a balanced multi-way search tree designed for the access characteristics of the disk. For an m-order B-tree, its main characteristics are: each node has at most m subtrees; except the root node, other nodes have at least m / 2 (rounded) subtrees; the root node has at least two subtrees, Except that the B tree only contains one node; all leaf nodes are on the same layer; the data on all nodes are arranged in order according to the key. The size of the nodes of the B-tree is usually the same as the block size of the disk. In this way, each time an element is searched, a node to which it belongs is prefetched t...

AI Technical Summary

Problems solved by technology

However, with the development of computer technology, the performance gap between memory and disk is getting bigger and bigger, and the performance of disk can no longer meet the needs of massive data analysis; data storage architecture has gradually centered on memory; and new storage devices, Such as PCM (Phase Change Random Access Memory) and 3D Xpoint, which have the characteristics of non-volatility of external memory, high integration and memory-level access speed, etc., which can make up for the volatile characteristics of existing DRAM (Dynamic Random Access Memory) , high energy consumption and limited capacity, the non-volatile memory (NVM, Non-Volatile Memory) based on these devices has become a hot spot for computer designers and researchers to improve the performance of storage systems

It is true that the design of the index structure for non-volatile memory also faces new challenges, such as non-volatile memory upgrading the original data persistence level from external memory to memory and its own write wear and other problems; generally speaking, memory The level of data persistence is to explicitly refresh the data in the CPU cache to the memory. For the B+ tree, due to the order of the elements on the leaf nodes, the insertion of the elements will cause the movement of the data, resulting in The overhead of a lot of CPU cache refreshing reduces the performance of the B+ tree; and in a multi-core system, the performance reduction of the single-threaded B+ tree operation will make it occupy the lock for too long, which will lead to more concurrent competition among multiple threads. Intense, further reducing the performance of the B+ tree and the utilization of system resources; the existing stand-alone can no longer meet the needs of large data storage, but in a distributed environment, because the processing of the B+ tree is server-centric, so The B+ tree design based on non-volatile memory faces the same challenges, such as the trade-off between data persistence and concurrency performance, and NVM write wear, etc.

Images