Introducing I-CASH, A New Disk IO Architecture

虽然存储容量和CPU处理能力在过去经历了快速增长，但硬盘驱动器（HDD）的数据带宽和访问时间的改进却没有跟上。因此，cpu和磁盘I/O之间的速度差距正在扩大。磁盘阵列可以提高总体I/O吞吐量，但由于涉及到机械操作，随机访问延迟仍然非常大。大缓冲区和深缓存层次结构可以改善延迟，但是由于磁盘存储级别的数据局域性差，访问时间的减少到目前为止非常有限。本提案旨在重新思考存储系统的基本架构，并尝试对基于磁盘的存储架构进行范式转换。该方法是建立一个新的存储架构，利用两种新兴的半导体技术：闪存SSD（固态磁盘）和GPU（图形处理单元）。新的磁盘I/O架构被称为I- cash: ssd和hdd的智能耦合阵列。SSD主要用于存储常读的“参考数据块”，以充分利用其高速随机读性能。HDD用于将当前I/O块与对应的参考块之间的压缩增量存储在SSD中，以便在线I/O操作时不会对SSD进行随机写操作。SSD和HDD由高速GPU控制，执行相似性检测、增量派生、将增量与参考块结合以及其他必要的功能，以实现存储与主机操作系统的接口。其理念是利用ssd的快速读取性能和gpu的高速计算，在很大程度上取代和替代HDD的机械操作，以实现比传统磁盘存储系统好几个数量级的I/O性能。与其在HDD上努力追赶处理器的性能，这已经被证明是困难的，如果不是不可能的话，建议的方法让存储系统驾驭多核处理器快速发展的浪潮，并通过以高速计算换取低访问延迟成为这种成功的一部分。预计拟议的项目将产生重大的广泛和变革性影响。1)数据中心的服务器运行着数十个和数百个虚拟机，这些虚拟机产生大量的I/ o，可以充分利用我们的新存储架构，潜在的数量级性能改进。2)这项研究将吸引研究生和本科生，以便他们为现实世界的需求做好准备。3)这项研究的成功将有助于罗德岛州和国家的经济发展。

While storage capacity and CPU processing power have experienced rapid growth in the past, improvement in data bandwidth and access times of hard disk drives (HDD) has not kept pace. As a result, the speed gap between CPUs and disk I/O is widening. Disk arrays can improve overall I/O throughput but random access latency is still very large because of mechanical operations involved. Large buffers and deep cache hierarchy can improve latency, but the access time reduction has been very limited so far because of poor data locality at the disk storage level.This proposal aims at rethinking the fundamental architecture of storage systems and makes an attempt at a paradigm shift of disk based storage architectures. The approach is to build a new storage architecture that exploits the two emerging semi-conductor technologies: flash memory SSD (solid state disks) and GPU (graphic processing unit). The new disk I/O architecture is referred to as I-CASH: Intelligently Coupled Array of SSDs and HDDs. The SSD is used to store mostly read "reference data blocks" to make best use of its high-speed random read performance. The HDD is used to store compressed delta between a current I/O block and its corresponding reference block in the SSD so that random writes are not performed on SSD during online I/O operations. The SSD and HDD are controlled by a high speed GPU that performs similarity detection, delta derivations, combining delta with reference blocks, and other necessary functions for interfacing the storage to the host OS. The idea is to leverage fast read performance of SSDs and the high speed computation of GPUs to replace and substitute, to a great extent, the mechanical operations of HDD to achieve I/O performance that is orders of magnitude better than traditional disk storage systems. Instead of working on HDD to catch up with processors' performance, which has been proven difficult if not impossible, the proposed approach lets storage systems ride the wave of the rapid advancement of multicore processors and be part of such success by trading high speed computation for low access latency. It is anticipated that the proposed project will have significantly broad and transformative impact. 1) Servers at data centers run tens and hundreds of virtual machines that generate large amount of I/Os that can take full advantage of our new storage architecture with potentially orders of magnitude performance improvement. 2) The research will engage both graduate and undergraduate students so that they are ready for the real world need. 3) The success of this research will help the economic development of the state of Rhode Island and the nation.