Flash Drives: Performance Study Based On BenchmarksChong Sun and Yupu ZhangComputer Sciences DepartmentUniversity of Wisconsin, MadisonAbstractFlash drives with nice properties, e.g. good random readperformance, has shown great potential to replace thecurrent dominant storage medium hard disks. However,there are not many literatures that have conducted com-prehensive study on the flash drive performance. In thispaper, we aim to design a systematic strategy to efficientlymeasure the performance parameters of flash drives. Wepropose several novel benchmarks to exploit the perfor-mance for us to better understand some interesting anderratic performance behaviors of flash drives.1 IntroductionThe largest obstacle that has been affecting I/O systems isno longer the storage volume nor the throughput, but theterrible random access performance of the hard disk drive.Due to the limit of mechanical rotating speed, high cost onposition seeking seems to be unavoidable for random diskaccess in the near future. The advent of flash memorypushes things a great step forward for its nice propertyof great random read performance. It is broadly recog-nized that flash memory can conduct uniformly randomread without “seek” time as in hard disk drives. Besides,flash memory generally has small size and less weightscompared with hard disks. Another attractive property isthat it is both shock resistant and immune to extreme tem-peratures.Though flash memory still has high cost/bit comparedwith hard disk drives, the volume of flash drive has beengradually increasing and corresponding the price has beendramatically decreasing. Recently, common laptops hasbegun to be configured with 32G or 64G flash based solidstate disks and we can even buy 2GB USB flash driveswith less than 20 dollars. More importantly, currentlymany companies with large data centers, e.g. Google, arespending a huge amount of money has on powering andcooling their hard disk drives. Flash memory drives gen-erally consume much less power and produce less heatthus saving much money.Flash drives bring us the promise of both good randomread and sequential access performance and has great po-tential of replacing hard disk drives in many places. How-ever, there are not many literatures on systematic studyof the real I/O performance on the flash drives. Eventhough some performance study of flash drives may havebeen conducted by the manufacturers, they are generallyconfined within the companies. Many performance pa-rameters on flash drives are not made public or hidden onpurpose by the manufacturers and it is difficult for us tounderstand the performance characteristics.In this paper, we aim to design a systematic strategyor benchmark to efficiently measure the performance pa-rameters of the flash drives. Initially, we conduct exper-iments with our designed benchmarks based on the ba-sic flash drives knowledge. Then we interpret the perfor-mance results that match our knowledge and redesign thebenchmarks to exploit the erratic I/O performance. We re-cursively refine our benchmarks until we collect enoughdrive parameters to systematically understand the I/O per-formance of the flash drives. Note that we generally needto provide assumptions for some parameters first and thendesign specific benchmarks to test according to our as-sumptions.Therefore, we may not thoroughly test all theuseful parameters, but we do systematically get the pa-rameters within our assumptions. We find some very in-teresting I/O performance behaviors, which can be effec-tively interpreted by the parameters we get with the de-signed benchmarks.We organize our paper as follows. First, we give somebackground knowledge about flash drives in Section 2.Then in Section 3, we design the prototype benchmarksto conduct some basic experimental study on the read andwrite performance. In Section 4, we propose several newbenchmarks to exploit performance and interpret some ir-ratic results. Section 5 presents several related work. Fi-nally, we give a conclusion and talk about the future workin Section 6.12 BackgroundA flash drive is usually composed of one or more flashmemory chips and a Flash Translation Layer (FTL),whose organization is illustrated in Figure 1.Figure 1: Flash Chip Organization2.1 Characteristics of Flash MemoryGenerally, there are two kinds of non-volatile flash memo-ries: NOR and NAND [10]. NAND is designed to satisfythe requirement of high capacity storage, while NOR isused to store small data and codes. So currently, almostall mass storage devices use NAND flash memory.A NAND flash memory chip consists of a fixed numberof blocks and each block has a fixed number of pages. De-pending on the manufacturer of the chip, each page couldbe 1KB, 2KB or 4KB and each block could contain 64,128 or 256 pages. Note that in the flash world, a page isthe unit of read and write, which is similar to a sector inhard disk or block in the I/O system. But a block, whichis confusing, is a very large unit here.A special property is that it cannot be overwritten di-rectly. Instead, any page in the flash can be rewrittenonly after the whole containing block is erased. Moreover,erase is much more time-comsuming compared with readand write operations (The time cost of each operation forone example is listed in Table 1). Besides, each block canonly tolerate certain times of erase, which is from 10,000to 100,000. All these limitations make the erase operationthe bottleneck of the whole flash memory.2.2 Flash Translation LayerDue to these limitations, a Flash Translation Layer is usedas a controller, emulating a hard disk, implementing theOperation Unit TimeRead 4KB Page 25usWrite 4KB Page 200us - 700usErase 256KB Block 1.2ms-2msTable 1: NAND Flash Operation Parameters[14]block functionality, and hiding the erase latency. There-fore, flash memory can be used as a normal block stor-age device in current operating systems without modifica-tions.FTL performs logic-to-physical address translation. Inorder to hide the erase latency, FTL usually redirects writerequests to a free block which is erased in advance. Thus,FTL must maintain internal logic-to-physical mapping in-formation and always keep it updated. Generally, thereare three types of translation schemes: page-mappingFTL [7], block-mapping FTL [5], log-block FTL [9],which combines both block-level mapping and page-levelmapping. A page-mapping scheme requires that the FTLmaintains a large page table such that any logical page ad-dress can