计算机英语 论文翻译例子(初级).doc

RETHINKING FLASH IN THE DATA CENTER反思闪存在数据中心寿浙威 非师 1271408006DEPLOYMENT OF FLASH MEMORY DEPENDS ON MAKING THE MOST OF ITS UNIQUE PROPERTIES INSTEAD OF TREATING IT AS A DROP-IN REPLACEMENT FOR EXISTING TECHNOLOGIES.闪存的未来应用取决于能否充分利用它的特性，而不要把它对待成为一个突然闯入并且来替代现有技术的。Over the past few years, computer systems of all types have started integrating flash memory. Initially, flashs small size, low power consumption, and physical durability made it a natural fit for mediaplayers and embedded devices. Lately, flashs rising density has won it a place in laptops and some desktop machines. 再过去的几年中，各种类型的计算机系统开始集成闪存。一开始，闪存体积小、低功耗以及物理上的持久特性使得它能很好的适合那些媒体播放器、嵌入设备。最近，闪存的高密度集成使它在笔记本电脑和台式电脑上也赢得了一些地位。Flash is now poised to make deep inroads into the data center. There, flash memorys high density, low power, and low-cost I/Os per second will drive its adoption and enable its application far beyond simple hard drive replacements. To date, however, many uses of flash have been hamstrung by a fundamental challenge of the technology: Flash is neither magnetic disk nor DRAM. It has its own performance advantages and quirks that system designers must address at several levels to best exploit it. 现在，闪存已经很自然的进入了数据中心。那些，闪存的高密度、低功耗和输入输出率的低消耗使得它的被采纳度和应用远远超越了简单的对硬盘驱动器的替代。然而，对于数据来说，大量对闪存的使用已经遭到了基础技术的质疑和阻碍：闪存既不是磁盘，也不是动态随机存取存储器。闪存拥有它自身的优点和怪癖，为了更好的利用它，系统设计师必须对它多重编址。Disk and DRAM replacement 代替磁盘和动态随机存取存储器So far, most proposed applications for flash in the data center have fallen into two categories.目前，很多对闪存在数据中心提出的应用可以分为两类。The first category is disk replacement. Quick access time and low power requirements make flash a compelling replacement for conventional disks, albeit at much lower density and higher cost. Accounting for servers and supporting infrastructure, solid-state disks (SSDs) consume roughly 10 times less energy when idle than disks. They deliver 2.6 times more bandwidth per watt and 3.2 times more bandwidth per dollar, 25 times more I/O operations per second (IOPS) per dollar, and 2,000 times more IOPS per watt .第一类就是代替磁盘。快速读取和小功率的要求使得闪存替代密度既低，消耗又高的传统磁盘引人注目，对于服务器的解释和基础设备的支持上，固态硬盘（SSDs）在系统闲置时的消耗时间大约是磁盘的十分之一，它们能传输2.6倍更多的带宽每瓦特，3.2倍更多的带宽每美元，25倍更多的输入输出操作每秒（IOPS）每美元、2000倍更多IOPS每瓦特。Flash sometimes also serves as a DRAM replacement. Density and (again) energy efficiency let flash compete with DRAM in applications where latency and bandwidth are less important. Flash consumes one-fourth the power of DRAM per byte at one-fifth the price.闪存有时候也充当动态随机存取存储器的作用。密度和能量效能使闪存能够在那些潜在因素和带宽不是很重要的应用领域中和DRAM竞争，因为闪存仅仅是DRAM每字节消耗的四分之一，同时价格是DRAM的五分之一。Flash memory will remain a contender for both roles for the foreseeable future, but additional opportunities and challenges are on the horizon. Technology scaling will continue to increase bit density for another 1 to 2 process generations, which will drive down costs. However, smaller flash cells are less reliable and less durable. In the past two years, lifetime program/erase cycle ratings for high-density flash devices have dropped from 10 thousand to five thousand cycles. Raw bit error rates have increased as well. Understanding how to apply this shifting technology in the changing landscape of datacenter computing requires careful design.在可预见的未来，闪存会是这两个方面冠军的争夺者。但是，其它的机会和挑战它们都在同一地平线线上。再用一、二代的过程，缩放技术会继续增加比特位的密度，降低成本。但是，更小的闪存单位会更不可靠，更不稳定。在过去的两年中，对于高密度的闪存设备，编码/擦除的生命周期率已经从10000降到了5000，生成的错误位率也随即增加。在数据计算不断改变的情况下，想要弄明白如何应用这个狡猾的技术需要细心地设计。Examples of how blithely applying SSDs to some datacenter applications can result in disappointing performance are easy to find. Our experience with a key-value store designed to hold huge numbers of small key-value pairs elicited nearly worst-case performance from SSDs. The FAWN-KV key-value storage system, developed as a part of Wimpy Nodes (FAWN) project, can handle more than 200 times more inserts per second than the traditional, non-flash-optimized Berkeley DB, on both older flash devices and modern SSDs.对于如何愉快的把SSDs应用到一些资料处理中心导致令人失望的例子处处可见。我们经历过，从SSDs用一个设计好的关键值来包含大量小的关键值组引出了几乎最糟糕的表现。 FAWN-KV关键值存储系统，作为Wimpy Nodes（FAWN）计划的一部分的发展，可以处理比传统，非闪存优化的Berkeley DB，无论是旧的闪存和现代的SSDs快每秒200倍的输入。 Flash translation layer闪存翻译层The flash translation layer (FTL) hides many of flashs remaining warts. It provides reliability and the abstraction of a uniform block address space. This is necessary since flash cannot do in-place updates, flash cells wear out after between five thousand and one million program/erase cycles, and even read operations can potentially corrupt data. Through heroic engineering and daunting complexity, the FTL masks these problems, but its performance impact can be significant. Intels Extreme SSDs have a read latency of 85 ms, but the flash chips the drive uses internally have a read latency of just 25 to 35 ms.闪存翻译层掩盖了它的一些缺点。它提供了可靠性和统一块空间的抽象，这是必须的，因为闪存不能再适当位置更新，闪存单元会在500000万次编码/擦除周期后损坏，甚至在读操作时会潜在地产生错误的数据。通过英勇的工程师和令人退缩的复杂性，FTL掩盖了这些问题，但是它所表现的影响是重大的。英特尔最先进的SSDs已经有一个85ms读取反应时间，但是在本地驱动闪存芯片只需要25到35ms读取反应时间。Flash clearly needs some kind of management layer to manage media errors and provide wear-leveling to maximize device lifetime. What interface should that layer present to the rest of the system, and where should it be implemented? Should the SSD give the application control over flash management, or should applications express their requirements to the SSD?很显然，闪存也需要一些管理层来管理媒体错误，以及提供平均磨损来提高设备的最大寿命。管理层需要给系统的其它部分呈现一个什么样的接口，并且哪里能使它生效？难道SSD应该对应用程序给予闪存的管理，或者应用程序向SSD发送它们的请求吗？Exposing a richer interface to the system might include providing multiple personalities to the system. These could include a disk personality (the current standard) as well as others that might expose flashs block and page structure while providing block-level wear-leveling. Another personality would expose the flash directly and require the system to manage wear-leveling and remapping as needed by that application. The interface could let the system partition the SSD and expose a different personality for each partition. For such a storage device, customization could occur in the kernel driver or even at the application level, letting applications map their storage needs directly and efficiently onto the flash. Aggregating management across multiple flash devices would let a large server (or even a server farm) make global optimizations based on error rate and performance variation, allowing it to extract the maximum performance from the flash array while meeting application-specific reliability targets.发掘一个对系统更加丰富的接口可能需要包含对系统提供的多重“个性”。这些可以包括磁盘特性（当前标准），以及其它可能会在提供块级别的磨损程度暴露闪存的块和页结构的特性。另一个特性是如果要求系统来管理磨损程度和应用程序对重测图的需要来直接暴露闪存。接口可以为系统提供给SSD分区和为每个分区发掘各自的特性。对每一个存储设备，在内核驱动甚至应用层可以被定制，让应用程序直接有效的在闪存上绘制它们的存储需要。横穿多重闪存设备的管理总计会使一个大型服务器（甚至是服务器群）在错误率和变化表现的基础上进行全局优化，允许它在遇到特殊应用程序的可靠目标上从闪存获得最大的表现。In the second approach, the FTL might provide an object store abstraction that lets applications indicate use patterns on a per-object basis. An application might designate one object for streaming, sequential writes and another for fast, random reads.在第二个分支上，FTL可能提供一个对象来存储来使应用软件以每个数据对象基础指明使用形式的抽象。一个应用程序可能标明了一个流对象，按次序的写和另一个更快，随机地读。Example applications应用例子We close with three examples of applications that could exploit a lower-level interface to flash.我们以三个开发一个闪存的低层接口的应用例子来结束。Bloom filters store set membership information, with a small chance of false positives, in a large bit array (for example, a flash page). To store an item, the filter computes k different hash values, and marks the bits addressed by those values. To test for membership, the filter applies the same hash functions and checks the k resulting locations. If all locations are marked, the item might be in the set.Bloom过滤器设置成员关系信息，在一个较少的不真实的正数的机会里，在一个大型位数组（举个例子，一个闪存页）。要存储一个项目，过滤器需要计算k次不同的哈希值，并且标记每个位用这些数编址来要测试成员关系。如果所有的位置都被标记 ，则这个项目可能在集合中。Flash device data sheets recommend programming each page of flash only once between erase operations, to manage the program disturb phenomenonthat is, a program to one page could corrupt data on another page. In practice, flash chips allow multiple programs to a page with the caveat that each program can only turn 1s (the erased state) into 0s (the programmed state). Data we collected on real devices demonstrates that the single-level cell flash devices found in highend SSDs can tolerate several hundred repeated programs before significant corruption occurs.This combination of characteristics makes these devices a perfect fit for implementing Bloom filters. To insert an element into a filter, the application programs the bit vector for the element onto the page. 闪存设备数据单子建议只为每一页闪存在擦操作直接编码一次，以此来管理程序扰动现象一个程序对应的那个页会破坏另一个页的数据。实际上，在附加上为每个程序仅仅只能从1（擦除状态）变为0（程序状态）后，闪存芯片允许多个程序同在一个页。我们在实际设备选择的数据表明单级闪存元设备加在SSDs后可以在重要的错误发生之前承受几百个重复的程序。这些特征的结合使得那些设备比较完美的符合Bloom过滤器的执行。为了在过滤器中插入一个元素，这个应用程序编码向量块来使元素进入页。The program operation performs an effective logical or with bits already programmed. To maintain data integrity, the application copies the Bloom filter to the next page every 1,000 insertions or so.程序操作执行一个有效逻辑或者块已经完成编码，为了维护数据的完整性，应用程序大约每1000插入后复制Bloom过滤器下一页。Other tricks are also possible. Write-once data encodings can allow applications to write multiple sets of arbitrary logical bits to a single page by encoding each pair of logical bits as three physical bits and ensuring that the second write only changes 1s to 0s. This technique can significantly improve flash lifetime and energy efficiency. The trade-off is that both of these uses violate the manufacturers guidelines. Whether that risk is worth additional lifetime or performance might best be decided on a perapplication basis.其它的技术也是可能的。数据一次编码可以允许应用程序写多组随意逻辑的小块到某个单页，通过编码每一对逻辑小块作为3个物理小块来保证二次写操作只从1改变到0。这个技术可以有效的提高闪存的寿命和能量效能。这两种的使用交换就是都违背了制造商的方针。无论磁盘值得增加额外的寿命或者表现可以最佳取决于应用程序的基础。A richer interface to flash would also enable less risky ventures by eliminating the one-size-fits-all FTL. Several important applications f