R&D: HyR-tree, Spatial Index for Hybrid Flash/3D XPoint Storage

Neural Computing and Applications has published an article written by Athanasios Fevgas, Data Structuring and Eng. Lab., Department of Electrical and Computer Engineering, University of Thessaly, 38221, Volos, Greece, Leonidas Akritidis, School of Science and Technology, International Hellenic University, Thessaloniki, Greece, Miltiadis Alamaniotis, Department of Electrical and Computer Engineering, University of Texas at San Antonio, San Antonio, USA, Panagiota Tsompanopoulou, Data Structuring and Eng. Lab., Department of Electrical and Computer Engineering, University of Thessaly, 38221, Volos, Greece, and Panayiotis Bozanis, School of Science and Technology, International Hellenic University, Thessaloniki, Greece.

Abstract: “Flash-based SSDs have become well established in the storage market, replacing magnetic disks in both enterprise and consumer computer systems. The performance characteristics of these new devices have prompted a considerable amount of research that aims at developing efficient data access methods. Early works attempted to reduce the expensive random writes, exploiting logging and batch write techniques, while more recent ones addressed query processing, taking advantage of the high internal parallelism of SSDs. 3D XPoint is a new nonvolatile memory technology that has emerged recently, featuring smaller access times and higher durability compared with flash. It is available both as block addressable secondary storage and as byte addressable persistent main memory. However, the high cost of 3D XPoint prevents, for the moment, its adoption in large scales. This renders hybrid storage systems utilizing NAND flash and 3D XPoint a sufficient alternative. In this work, we propose HyR-tree, a hybrid variant of R-tree that persists a part of the tree in the high performing 3D XPoint storage. HyR-tree identifies repeated access pattern to the data and uses these patterns to locate the most important nodes. The importance of a node is determined by the performance gain that derives from its placement within a 3D XPoint-based device. We experimentally evaluated HyR-tree using real devices and four different datasets. The acquired results show that our proposal achieves significant performance gains up to 40% in tree construction and up to 56% in range queries.“