A Common Storage Engine for Modern Memory and Storage Hierarchies

点击翻译按钮获取中文摘要

Scientific research is increasingly driven by data-intensive problems. As the complexity of studied problems is rising, so does their need for high data throughput and capacity. The globally produced data volume doubles approximately every two years, leading to an exponential data deluge. This deluge then directly challenges database management systems and file systems, which provide the foundation for efficient data analysis and management. These systems use different memory and storage devices, which were traditionally divided into primary, secondary and tertiary memory. However, with the introduction of the disruptive technology of non-volatile RAM (NVRAM), these classes started to merge into one another leading to heterogeneous storage architectures, where each storage device has highly different performance characteristics (e.g., persistence, storage capacity, latency). Hence, a major challenge is how to exploit the specific characteristics of memory devices.To this end, SMASH will investigate the benefits of a common storage engine that manages a heterogeneous storage landscape, including traditional storage devices and non-volatile memory technologies. The core for this storage engine will be B-epsilon-trees, as they can be used to efficiently exploit these different devices. Furthermore, data placement and migration strategies will be investigated to minimize the overhead caused by transferring data between different devices. Eliminating the need for volatile caches will allow data consistency guarantees to be improved. From the application side, the storage engine will offer key-value and object interfaces that can be used for a wide range of use cases, such as high-performance computing (HPC) and database management systems. Moreover, due to the widening gap between the performance of computing and storage devices as well as their stagnating access performance, data reduction techniques are in high demand to reduce the bandwidth requirements when storing and retrieving data. We will, therefore, conduct research regarding data transformations in general and the possibilities of external and accelerated transformations. As part of SMASH, we will provide a prototypical standalone software library to be used by third-party projects. Common HPC workflows will be supported through an integration of SMASH into the existing JULEA storage framework, while database systems can use the interface of SMASH directly whenever data is stored or accessed.The consortium will be supported by Intel Corporation, especially concerning the hardware acceleration of data transformations and collaborate actively with other project within the SPP 2377.