数据存储处理器——将智能用于存储

Data-storage Processing Unit – Placing Intelligence in Storage

  演讲人：

杨亚飞

深圳大普微电子科技有限公司CEO

Yafei Yang

CEO

Shenzhen DAPU Microelectronics Co., Ltd

演讲人简历：

杨亚飞博士，微电子与信息安全领域专家，在半导体领域拥有多年开发和管理经验。2008年5月至2016年3月曾在美国高通公司总部工作，负责高通产品安全战略，包括智能手机芯片、汽车芯片、网络芯片、物联网芯片和服务器芯片等。杨博士2016年4月回国创办了深圳大普微电子科技有限公司，致力于打造世界首创的数据存储处理器芯片，项目荣获深圳市“孔雀计划”和4000万元研发资助。杨博士毕业于美国罗德岛大学，拥有7项美国专利，2010年曾受邀担任美国白宫国家宽带计划（BTOP）项目评审专家，并受到时任美国商务部部长骆家辉先生和副部长Lawrence Strickling的感谢和表彰。

Dr. Yafei Yang is an expert in chipset development and information security, with many years of research and management experience in semiconductor industry. From May 2008 to March 2016, he worked at Qualcomm Inc., responsible for Qualcomm product security strategy, including smartphone chipsets, automotive chipsets, networking chipsets, IoT chipsets and server chipsets. In April 2016, Dr. Yang Co-founded Shenzhen DAPU Microelectronics Co. Ltd and received initial funding from Shenzhen Peacock Plan, a very prestigious fund for high tech startups. Dr. Yang graduated from Univ. of Rhode Island and holds 7 US patents. He was invited to perform expert reviews on the technical proposals submitted under BTOP (White House Broadband Technology Opportunities Program) in 2010. To honor Dr. Yang's outstanding contribution to US economic recovery, US Secretary of Commerce (Gary Locke) and Assistant Secretary (Lawrence Strickling) wrote appreciation letters to him.

报告摘要：

现有的SSD和HDD等存储设备都用传统的MCU和固件来管理和控制存储数据。随着新型存储介质技术如闪存、PCM、RRAM、MRAM等不断涌现，管理和控制数据存储变得越来越复杂。另一方面，在大数据时代，结构化数据和非结构化数据的指数型爆发增长对存储管理提出了空前的挑战，大数据应用的多样性也对存储的I/O性能、安全性和可靠性提出了更高的要求。因此，现在已经到了一个新的历史时刻，需要我们重新考虑存储控制器的架构，使得更好的适应大数据和新存储介质的发展。

为此，我们首次提出一个新概念——数据存储处理器（Data-storage Processing Unit，简称DPU）。DPU是一种专门针对大数据、云计算特性而定制设计，用于管理、控制和分析大数据并对数据进行分类。DPU把智能带入到SSD存储设备中，通过机器学习算法来不断学习和优化I/O性能，提高数据可靠性、安全性和可用性，从而针对各种不同的大数据应用达到最优的管理和控制性能。

Traditionally, storage devices such as HDD and SSD are managed and controlled by standard MCUs and firmware running on them. As new storage technologies emerge such as flash memory, PCM, Memristor, and MRAM, management and control of storage devices have become more and more complex. On the other hand, data that are stored in storage devices grow exponentially encompassing all kinds of data including structured and unstructured data, which further complicates data storage management. In addition, variety of big data applications demands high I/O performance and high data reliability/availability, placing stringent requirements on storage devices. Therefore, it has come to a historical point where we should reconsider storage controller’s architecture to keep pace with the explosive growth of big data and fast emergence of new storage media technologies.