Font Size: a A A

Research On Evaluation Model And Optimization Of Limited Program/Erase Cycles (P/Es) In NAND Flash-based SSDs

Posted on:2015-11-12Degree:DoctorType:Dissertation
Country:ChinaCandidate:H SunFull Text:PDF
GTID:1228330428465748Subject:Computer system architecture
Abstract/Summary:PDF Full Text Request
NAND Flash-based solid state disks (i.e., SSDs) are characterized with the limited number of program and erase cycles (i.e., P/Es), which are dependent on the P/Es of NAND Flash inside SSDs. There are the following shortcomings observed in NAND Flash:the limited number of P/Es and restrictive operations (i.e., sequential write, out-of-place update and erase-before-write). As long as scaling of NAND Flash memory process technology, the number of P/Es reduce,which degrade NAND Flash endurance and SSD lifetime. It is important to propose accurate models to evaluation the utilization of P/Es in NAND Flash and to design novel technologies to improve the utilization of P/Es.An accurate and practical write amplification model is proposed. Write amplification is used to evaluate the effect of extra program operations on the limited number of pro-gram/erase cycles (i.e., P/Es) in NAND Flash. The lack of an accurate way to measure write amplification for SSDs motivates me to propose a novel measuring method called RB-Explorer at the SSD level rather than the NAND Flash level. RB-Explorer incorporates a Ready/Busy (R/B) signal of one of the NAND Flashes in an SSD in a proposed write am-plification model for SSDs with four full-parallelism levels (i.e., the channel, chip, die, and plane levels). The duration of a low level of R/B varies with different operations in NAND Flash. RB-Explorer scan the output level of an R/B signal to calculate the number of the low levels of the R/B signal for page programs. RB-Explorer takes two steps toward measur-ing write amplification. First, RB-Explorer quantifies the number of page programs using the low R/B signal level. Second, RB-Explorer measures data volume written to NAND Flashes by considering parallelism at four levels. Data volume written to a die in a NAND Flash is obtained as a product of the number Np of programs and page size Pa. Given the number Nchannel of channels, the number Nchip of chips per channel, and the number Ndie of dies per chip, one can obtain the data volume written to NAND Flashes as a product of Np,Pa, Ndie, Nchip, and Nchannel.Write amplification is a ratio of data volume written to the NAND Flashes by an SSD controller to data volume written from a host. The former one can be measured practically and accurately by RB-Explorer, the latter one can be obtained by users. A real-world SSD (i.e., SSD-v) is implemented and employ a fine-tuned SSD sim-ulator (i.e., SSDsim) to validate the accuracy of RB-Explorer. Experimental results show that RB-Explorer improves on the accuracy of SSDsim in most tested cases. A series of measurements using micro-benchmarks and I/O traces are conducted to demonstrate how RB-Explorer may be applied to investigate SSDs.A comprehensive Behavior amplification model (i.e., Bamp) is proposed to explore all behaviors (i.e., read, program, and erase) in SSDs. This model is on the basic of SSD perfor-mance and energy consumption. With NAND flash process development, the P/Es in NAND Flash decrease. The studies revealed that the erase and read operations can degrade P/Es. Overhead caused by read and erase operations is not taken into account by write amplifi-cation. This typical overhead gives rise to bad impacts on data reliability in NAND Flash space as long as scaling of NAND flash memory process technology. Bamp model is pro-pose to evaluate overhead of program, erase, and read operations on NAND Flash endurance and SSD lifetime. Bamp is the ratio of actual energy consumption of all behaviors to user-written-data-based energy consumption. The value of Bamp is can be obtained in the term of SSD performance and SSD energy consumption when user data is written to devices. Given a write-dominated workload condition, the user-written data can be obtained according to SSD performance, while user-written-data-program-and overhead-operation-based energy consumption can be obtained by measuring SSD energy consumption. Besides, a method is employed to evaluate SSD performance per energy consumption and to find the SSD which provides high performance at low energy consumption under a combination of file system and I/O scheduler. This method is to reduce the effect on confirm the accuracy and credibil-ity of Bamp model.Experimental results reveal that the method to studying Bamp by SSD performance and energy consumption is accurate, and; Comparing with write amplification, Bamp can comprehensively evaluate effects of extra operations on P/Es in NAND Flash and SSD lifetime.An effective log-space allocation strategy called distributed log space (i.e.,DLSpace) is proposed. Restricted operations (i.e., sequential write, out-of-place update and erase-before-write) can trigger write amplification and block erase operations, which degrade the limited number of P/Es and shorten SSD lifetime due to adverse impacts on NAND Flash endurance. The traditional log space allocation strategy (i.e., TLSpace) hardly handles the increment of write amplification and the number of erase operations. DLSpace is designed to reduce write amplification and the number of erase operations. The DLSpace-based log space is divided into log block space and log page space. Log page space in DLSpace is composed of the portion of data pages in a data block and only performs as buffer for updated pages in this data block. DLSpace strategy gives a high priority to log page space by utilizing log pages before any log block is used. Log page space buffers the page-update operations to delay use of log blocks for postponing garbage collection. DLSpace fully utilizes pages in data and log blocks for reducing erasures of blocks with free pages. Meanwhile, DLSpace decreases write amplification and block-erase operations by reducing excessive valid page-rewrite and delaying garbage collection under workload dominated with random writes. Then, the utilization of P/Es in NAND Flash is enhanced and SSD lifetime is improved. Experimental results reveal that compared with the existing TLSpace, DLSpace-based SSDs are characterized with the lower write amplification, the longer delay time of garbage collection, and the fewer number of erases operations, which can be further improved by increasing the number of log page space. DLSpace is conducive to slowing down the P/E degradation and prolonging SSD lifetime.
Keywords/Search Tags:NAND Flash, Solid State Disk, P/Es, Write Amplification, Behavior Amplification, Log space allocation strategy
PDF Full Text Request
Related items