Reinforcement learning-driven address mapping and caching for flash-based remote sensing image processing

Abstract

Flash memory is featured with salient advantages over conventional hard disks for massive data storage and efficient on-board data processing. A flash translation layer (FTL) is a critical component for flash-based storage devices to handle particular technical constraints of flash. It is desirable to use flash memory for the storage of massive remote sensing images and support on-board remote sensing data processing applications, which typically require high I/O performance and hence call for advanced FTL design and implementations. In this paper, we introduce our efforts in developing a reinforcement learning driven page-level mapping and caching scheme (named Q-FTL) that is adaptive and responsive to ever-changing I/O streams of on-board remote sensing image processing operations. The adaptability and responsiveness are achieved by the separation of large and small I/O requests, an integrated weighting scheme to measure access costs of cached translation pages, and a reinforcement learning driven cache replacement algorithm. We demonstrate the efficiency of the proposed approach using actual I/O traces generated from on-board remote sensing image processing applications. Experimental results show that Q-FTL improves over several current state-of-the-art FTLs by a large margin and even achieves competitive performance close to an idealized pure page mapping FTL in some cases.

Introduction

The fast advancement of earth observation technologies creates an ever-increasing gap between data acquisition and data processing capabilities. Massive volumes of remote sensing image data are being collected at an unprecedented pace. Typically, these image data cannot be used before being downlinked to ground stations, in which image data are processed and distributed to end users. This traditional image data processing procedure poses challenges for the efficient transmission of massive remote sensing image data. An alternative solution is to transfer image processing capabilities from ground to sensor platforms and to perform on-orbit (or in-flight) image processing [1], [2], [3], [4]. Under such circumstances, an appropriate storage solution is urgently needed to underpin high performance I/O. Over the last decade, we observe a wide embrace of NAND flash memory as a persistent storage media in both portable devices (e.g., digital cameras, tablets, and mobile phones) and enterprise storage solutions [5], [6], [7]. Flash memory has salient advantages over conventional hard disks, such as fast random access rate and shock resistance. These merits make it a suitable media for on-board remote sensing data storage. Other than these advantages, flash memory has some unique properties that must be considered in the development of flash-based data storage and management technologies. These particularities include: I/O asymmetry between read and write, the erase-before-write constraint, and limited life expectancy. Consequently, flash needs to perform out-of-place update to avoid degraded performance caused by erase operations.

A flash translation layer (FTL) can be implemented and deployed in flash controllers to handle above-mentioned technical constraints of flash [8]. The primary function of an FTL is to perform logical-to-physical address translation to emulate traditional block devices and provide backwards compatibility for high-level applications running on flash memory. T

According to mapping granularity, three types of address mapping methods can be identified: page-level mapping [9], block-level mapping [10], [11], and hybrid mapping [12], [13]. A page-level mapping is highly efficient and flexible since address translation is performed at a finer granularity of page level. This study adopts a page-level mapping approach and uses advanced cache management mechanisms to exploit SRAM economically.

Some efforts have been made to develop specific flash-aware index structures for geospatial data [14], [15]. However, little effort has been made to develop flash-based storage solutions dedicated for remote sensing data processing applications, which often manifest unique I/O characteristics and require high I/O throughput rates with limited SRAM size. Zhang et al. [16] proposed a probability-based cache management approach for remote sensing image processing tasks. Nevertheless, this approach is largely static and its adaptability can be improved by machine learning technologies. Fig. 1 presents visualizations of typical non-spatial I/O traces and remote sensing processing I/O traces. Compared to generic non-spatial I/O traces, remote sensing traces exhibit strong spatial locality and correlated access patterns. Large-sized sequential I/O requests in remote sensing traces are multiplexed with small-sized random reads, which are subject to change over time. These dynamic I/O behaviors cannot be well handled by existing cache management schemes. The lack of adaptive data-dependent cache management methods in existing FTLs may lead to decreased hit ratios and inferior I/O performance for remote sensing applications. To address this issue, we need a data-driven modeling approach to separate I/O request with distinct patterns, to provide accurate measurements on access costs on cached data, and to enable adaptive cache replacement with acceptable overhead.

In this paper, we propose a page-level mapping and caching scheme (named Q-FLT) that is adaptive and responsive to ever-changing I/O streams of on-board image process operations. The adaptability and responsiveness are achieved by the separation of large and small I/O requests, an integrated weighting scheme to measure access costs, and a reinforcement learning driven cache replacement algorithm.