Section: New Results

Flash-Based Data Management

Participants : Nicolas Anciaux, Matias Bjørling, Philippe Bonnet, Luc Bouganim [correspondent] , Niv Dayan, Saliha Lallali, Philippe Pucheral, Iulian Sandu Popa.

There is a long tradition of work around the understanding and optimization of NAND Flash memory in the team (e.g., [7] , [9] ). Current work in this area covers the optimization of SSD use in DBMS engines and the design of Flash-based indexing techniques for textual and spatio-temporal data. These works on Flash-Based indexing complete the work initiated in the last years on the storage and indexing engine of PlugDB (not repeated in this report but the interested reader is referred to a DAPD'14 journal publication detailing these techniques [14] ).

Flash storage optimization. Solid State Drives (SSDs), based on flash chips, are now the secondary storage of choice for data intensive applications. Database systems can now rely on high performance SSDs to store log, indexes and data either on servers or in the cloud. While SSDs provide increasingly high performance out of the box, maintaining high throughput and low latency as the utilization of SSDs increases and despite abrupt changes in the workload remains a challenge. This question is central for database designers and administrators, cloud service providers, and SSD constructors. The answer depends on write-amplification, i.e., garbage collection overhead. More specifically, the answer depends on how write-amplification evolves in time. We derived a mathematical expression that relates over provisioning to write-amplification. We introduced a new block manager, called Wolf, or WOrkload Leveler for Flash. Wolf is able to detect and quickly adapt to changes in workload by pro-actively reallocating over-provisioned space among the groups based on their changing needs. It adapts better to stable workloads by measuring the update frequencies of groups instead of making assumptions about them. It uses a novel near-optimal closed-form expression to allocate over-provisioned space to groups.

Flash-based keyword indexing. As smart objects gain the capacity to acquire, store and process large volumes of data, new services emerge. However, the smart objects have to be endowed with typical data management capabilities to enable all these services. In this work, we revisit the traditional problem of information retrieval queries over large collections of files in an embedded context. A file can be any form of document, picture or data stream associated with a set of terms. A query can be any form of keyword search using a ranking function (e.g., TF-IDF) identifying the top-k most relevant files. The proposed search engine can be used in sensors to search for relevant objects in their surroundings, in cameras to search pictures by using tags, in personal smart dongles to secure the querying of documents and files hosted in an untrusted Cloud or in smart meters to perform analytic tasks (i.e., top-k queries) over sets of events (i.e., terms) captured during time windows (i.e., files) [21] . Designing such embedded search engine is however challenging due to a combination of severe and conflicting hardware constraints (e.g., a tiny RAM combined with a NAND Flash persistent storage badly adapted to random fine-grain updates). To tackle this challenge, we introduce three design principles, namely Write-Once Partitioning, Linear Pipelining and Background Linear Merging, and show how they can be combined to produce an embedded search engine reconciling high insert/delete/update rate and query scalability. We have implemented our search engine on a development board having a hardware configuration representative for smart objects. The experimental results demonstrate the scalability of the approach and its superiority compared to state of the art methods [28] . This work is part of Saliha Lallali’s Ph.D. thesis.

Flash-based spatio-temporal indexing. The convergence of mobile computing, wireless communications and sensors has raised the development of many applications exploiting a massive flow of spatio-temporal data such as location-based services, participatory sensing, or traffic management [15] . Among the most active research topics in this area is the spatio-temporal data indexing. Nevertheless, since a few years a new fundamental parameter has made its entry on the database scene: the NAND flash storage. However, the peculiar characteristics of flash memory require redesigning the existing data storage and indexing techniques that were devised for magnetic hard-disks. In this study we propose TRIFL, an efficient and generic TRajectory Index for FLash. TRIFL is designed around the key requirements of trajectory indexing and flash storage. TRIFL is generic in the sense that it is efficient for both simple flash storage devices such as the SD cards and more powerful devices such as the solid state drives. In addition, TRIFL is supplied with an online self-tuning algorithm that allows adapting the index structure to the workload and the technical specifications of the flash storage device to maximize the index performance. Moreover, TRIFL achieves good performance with relatively low memory requirements, which makes the index appropriate for many application scenarios. The experimental evaluation shows that TRIFL outperforms the representative indexing methods on magnetic disks and flash disks. This work is part of Dai-Hai Ton That Ph.D. thesis, co-supervised by Iulian Sandu Popa.