还记得上一篇博文中我提到过一个特别的咒语——大家不要再把NVM当作快速存储设备，而要将它看成是（慢速）存储设备！看起来，我并不是唯一持有这个想法的人，在由UCSD的两个研究小组组织的NVM工作室的讨论中，处处体现出了这一假设的精神。
作为今年活动的白金赞助商，PMC一直认为该工作室在技术上与闪存峰会相对应。此项活动在规模上远比FMS要小，但技术性更强，由工业界及学术界专家组成的讲者阵营提供了良好的平衡视角。我已连续三年参与此项活动，感到在此总能及时把握行业专家的研究动态，并接触到各色研究生及博士生等（可能受邀加入PMC的候选人）。
今年，参会人数达到了212人，创下了历史新纪录，其中，52%来自工业界，45%来自学术界，其余3%则来自其他领域（即政府等）。PMC在活动中张贴了两张大幅展板，一张描述了如何在NVMe SSD中控制QoS指标，另一张则陈列了PMC在开展的开源Donard计划的最新进展。此处 提供了更多关于Donard的信息。会上有许多精彩的主题演讲、技术论坛及展板展示等等。在此，我想分享一些我认为引人注目的亮点。请注意：我并未能聆听到所有的讨论，所以这只是个人参与活动的一些观感：
§ 主题演讲：今年共有三场主题演讲。Samsung的Bob Brennan和Seagate的Phil Brace的演讲均比较高层，焦点放在HG-NVM出现后将会带来的存储增长与变化。Intel 的Andy Rudoff则采取不同的视角，演讲更加技术性，广受听众赞赏。其演讲深入阐述了Intel的ISA和Linux是如何改变以适应持久性主存。三个演讲的内容均十分丰富详实。
§ 纠错：该工作室一直很强调ECC方面的工作。今年，一次写入存储（WOM）代码的讨论仍在继续。虽然，从理论上来讲，我很喜欢WOW代码，但对其在SSD中的实际使用价值表示怀疑，原因是该代码需要牺牲容量来换取耐久度，因此必然会提升高含金量的$/GB指标。话说回来，如果有人能证明我的谬误，我会很乐意接受指教。在会上，也看到针对NVM对LDPC代码（我挚爱的话题）做出的若干改进。针对NG-NVM作的ECC鲜有出现，这点很令人遗憾，因为我认为，此类存储预计会提供的低延迟特性将为ECC带来新的挑战，我很希望能看到在该领域的更多耕耘。
§ 内存设备：内存公司一直在讲要推出NV-NVM。然而，实际数据还是不太明确，而今市场上只有Everspin量产了DDR3 ST-MRAM产品（目前密度尚低）。NAND从2D向3D的过渡确实让 NAND扩展生机复燃，其他替代技术眼下看起来也不再象此前那么有吸引力。我们仍然继续梦想着有一种持久性的内存技术，从访问时间、成本及耐久度等指标来看，性能介乎DRAM与NAND之间。
§ SSD技术的发展：我目前感兴趣的一个领域是基于主机的FTL SSD。此类设备将眼下放在SSD上的“智能”（如垃圾回收）移至主机之上。确然无疑地，此类设备的大规模部署即是实证，显示着NAND上的成本支出能由此得到大幅度缩减。百度在此提供了一篇相关的优秀论文。由哥本哈根信息科技大学（IT University of Copenhagen）PhilipeBonnet领导的一个研究小组也在此领域开展了有效的工作，并通过OpenChannelSSD项目提供了开源代码。另外一个话题是“数据感知型SSD”,关注点在于如何使用（从应用层提供的）提示信息来告知（感知型）SSD在何处及如何存放数据，以使性能及耐久度达到最高。
§ 应用：应用需要如何改变以支持NVM，这个领域受到了广泛的关注。单单改变操作系统是不够的。比如，近来已经看到Hadoop 及其NoSQL通过诸如Aerospike 改进了对NVM的支持。前Fusion-IO（而今的SanDisk）在会上作了一场有关如何利用NVM来改善SQL压缩性能的精彩演讲。
总而言之，今年的NVM工作组成绩卓著。研究NVM技术的团队人才济济，大家为NVM进行的操作系统及应用层面的优化日益增多。我相信，随后的12个月内，还将会取得非常多重要的进展，无疑将为下一届NVM工作室的成功打下坚实的基础！
Recall from my last blog that I have a mantra – Let’s stop thinking aboutNVM as fast storage and start thinking about it as (slow) memory! Well it seemsI am not alone in that thinking, as this premise was very well represented atthe NVM Workshop organized by two of the research groups at UCSD.
PMC had the pleasure of being aPlatinum Sponsor at this year’s event and I always consider this to be thetechnical counter-point to Flash Memory Summit. The event is a lot smaller thanFMS but more technically orientated with a nice mix of industrial and academicspeakers. I have been attending for three years now and always find it a greatplace to catch up on people’s research and meet graduate students (who might bepersuaded tocome work for PMC).
This year a record 212 peopleregistered with 52% being industrial, 45% academic and 3% other (i.e.government). PMC had two posters at the event. One of these looked atcontrolling QoS metrics in NVMeSSDs, the other looked at some recent developments inour open-source Donard project. You canread more about Donard here. There were some great keynotes, technicalsessions and posters and I wanted to call out some of my highlights. Pleasenote: I did not make it to all the talks, so this is just my view of whatI saw at the event:
§ Keynotes: Thisyear we had three keynotes. Both Bob Brennan (Samsung) and Phil Brace (Seagate)were a bit more high level with a focus on storage growth and changes that willoccur as HG-NVM come on board. Andy Rudoff (Intel),took a different tack andwent technical, which I think this audience appreciated, delving into howIntel’s ISA and Linux are changing to accommodate persistent main memory. Allthree talks were very informative but Andy’s definitely won the Geek award ;-).
§ ErrorCorrection: The workshop always has a strong ECC component.This year we saw a continued discussion on Write Once Memory (WOM) codes. WhileI like WOM codes from a theory point of view, I struggle to see how they can beused in practical SSDs given that they sacrifice capacity for endurance, andtherefore increase that golden metric of $/GB. However, I am quite happy to beproven wrong. We also saw some improvements to LDPC codes (a topic very close to my heart), for NVM. I did not see much onECC for NG-NVM, which is a shame as I think the low latencies predicted forthose memory types will place new pressures on ECCs, and this an area I wouldlike to see more research in.
§ MemoryDevices: We continue to see memory companies discussingthe roll-out of NG-NVM. However, real data is still somewhat lacking with onlyEverspin in production today with their DDR3 ST-MRAM products, (and these arelow density for now). The NAND transition from 2D to 3D has definitely put somelife back into NAND scaling, and replacement technologies are probably not asappealing as they were before. We still continue to dream of a persistentmemory technology that lies somewhere between DRAM and NAND in terms of accesstimes, cost and endurance.
§ Advancesin SSDs: One area I am interested in right now isHost-Based FTL SSDs. In these devices some of the “smarts” that currently livein the SSD itself (e.g. garbage collection),is moved to this host. In certain,large scale, deployments there is some evidence to suggest significant savingsin NAND costs can be made. There is a good paper on this topic by Baiduhere. A teamat IT University of Copenhagen led by Phillipe Bonnet is doing some interestingwork in this space and going open-source with it via theOpenChannelSSD project. Another topic was “aware-SSDs” and how we can usehinting to tell the SSD how and where to place the data to maximize performanceand endurance.
§ Applications: An areathat received a lot of attention was how applications need to change to supportNVM. Changes to the OS are not always enough. For example, we have recentlyseen improved NVM support inHadoop and in NoSQL via things likeAerospike. FormerFusion-IO (now SanDisk), gave a nice presentation on using NVM to improve theperformance of SQL compression.
All-in-all it was another great year for the NVMWorkshop. There are a lot of very clever people working on NVM technologies andwe are starting to spend more and more time optimizing OS’s and applicationsfor NVM. I am sure there will be plenty of great developments over thenext 12 months that will ensure the next NVM Workshop continues to be moreinteresting and successful year-over-year!
原文地址：http://blog.csdn.net/pmc/article/details/45333773
