[IEEE Telecommunication Systems (MASCOTS) - Baltimore, MD, USA (2008.09.8-2008.09.10)] 2008 IEEE...

2
Write Buffer-aware Address Mapping for NAND Flash Memory Devices* Sungmin Park, Hoyoung Jung, Hyoki Shim, Sooyong Kangt Jaehyuk Cha Division ofInformation and Communications, Hanyang Univ., Seoul, 133-791, Korea {syrilo, horong, dahlia, sykang, chajh}@hanyang.ac.kr Abstract merge the original data block and the corresponding page cluster in the write buffer at the time when the page cluster By using small-sized, next-generation NVRAM (such as is evicted from the buffer, neither the sector-mapping table MRAM, FeRAM and PRAM) as a write buffer, we can im- nor the smart copy operation are needed. Also, page in- prove the overallperformance of the NANDflash memory- validation is not needed since updated pages are accessed based storage systems. However, traditional address map- from the write buffer instead of the flash memory. In this ping algorithms in Flash Translation Layer (FTL) software way, page clustering in the write buffer can dramatically were designed without any consideration of the existence simplify the FTL algorithm. of write buffer In this paper, we propose a novel write buffer-aware flash translation layer algorithm, optimistic 2. Optimistic FTL FTL, which is designed to harmonize well with NVRAM write buffers. Simulation results show that the proposed optiisti FT outerfoms revius lg bock-asedFTLIn this section, we propose a novel write buffer-aware alortitmis. FTL outperforms previous log block-based FTL FTL, Optimistic FTL, which exploits the characteristics of the block level write buffer management policies. It as- sumes that the write buffer is used for page clustering and that all pages in the victim cluster are passed to the FTL to 1. Introduction be written to the flash memory, in the order of page number. The term 'Optimistic' is based on this assumption. Since previous FTL algorithms did not consider the ex- The Optimistic FTL is much simpler than the previous istence of the NVRAM write buffer, their design policy log block based FTL algorithms. It has only one log block 'efficiently accommodate different write patterns (ex. ran- and maintains only block mapping table in the main mem- dom small writes or sequential large writes) to the flash ory. While the BAST and FAST algorithms maintain not memory' added large complexity to the FTL. For example, only a block mapping table but also a sector mapping table BAST [1] and FAST [2] algorithms use log blocks to cope for the log block, Optimistic FTL does not maintain a sec- with random small writes. A page is not updated but is in- tor mapping table since it does not use sector mapping for validated and the new version of the page is written in the the log block. Also, instead of three types of merge opera- log block. tions - switch, smart copy and switch copy, Optimistic FTL Using an NVRAM write buffer for page clustering has only one type of merge operation: Instant merge. Fig- (where pages in the buffer are clustered according to the ure 1 shows the instant merge operation. Assume that there block number in the flash memory and a page cluster is se- are eight pages in a block and currently all pages in a data lected as a victim for replacement), small random writes block are filled with valid data pages. The write request can be transformed into sequential writes to the flash mem- queue in the figure shows the sequence of victim pages that ory [3]. If we can delay writing a page cluster until the should be written to the flash memory. We can see that cluster become sufficiently large, it is not necessary to de- pages 3, 4 and 7, which belong to the same data block in lay erasing data blocks in the flash memory because many the flash memory, are selected together as victim pages, in of the pages in the data block are already obsolete and the order of page number, according to the block based write number of valid pages in the original data block, which buffer management policy. When a write request to page should be copied to the new data block, is small. If we 3 arrives, Optimistic FTL copies pages 1 and 2 from data *This work was supported by grant No. RO1-2007-OOO-20649-O from block to the log block ( I, O) and then writes page 3 in the Basic Research Program of the KOSEF. the buffer to the log block((c)). On receiving a write re- tCorresponding author quest to page 4, it writes the page into the log block at

Transcript of [IEEE Telecommunication Systems (MASCOTS) - Baltimore, MD, USA (2008.09.8-2008.09.10)] 2008 IEEE...

Page 1: [IEEE Telecommunication Systems (MASCOTS) - Baltimore, MD, USA (2008.09.8-2008.09.10)] 2008 IEEE International Symposium on Modeling, Analysis and Simulation of Computers and Telecommunication

Write Buffer-aware Address Mapping for NAND Flash Memory Devices*

Sungmin Park, Hoyoung Jung, Hyoki Shim, Sooyong Kangt Jaehyuk ChaDivision ofInformation and Communications, Hanyang Univ., Seoul, 133-791, Korea

{syrilo, horong, dahlia, sykang, chajh}@hanyang.ac.kr

Abstract merge the original data block and the corresponding pagecluster in the write buffer at the time when the page cluster

By using small-sized, next-generation NVRAM (such as is evicted from the buffer, neither the sector-mapping tableMRAM, FeRAM and PRAM) as a write buffer, we can im- nor the smart copy operation are needed. Also, page in-prove the overallperformance ofthe NANDflash memory- validation is not needed since updated pages are accessedbased storage systems. However, traditional address map- from the write buffer instead of the flash memory. In thisping algorithms in Flash Translation Layer (FTL) software way, page clustering in the write buffer can dramaticallywere designed without any consideration of the existence simplify the FTL algorithm.of write buffer In this paper, we propose a novel writebuffer-aware flash translation layer algorithm, optimistic 2. Optimistic FTLFTL, which is designed to harmonize well with NVRAMwrite buffers. Simulation results show that the proposed

optiisti FTouterfoms revius lg bock-asedFTLIn this section, we propose a novel write buffer-awarealortitmis.FTL outperforms previous log block-basedFTL

FTL, Optimistic FTL, which exploits the characteristics ofthe block level write buffer management policies. It as-sumes that the write buffer is used for page clustering andthat all pages in the victim cluster are passed to the FTL to

1. Introduction be written to the flash memory, in the order ofpage number.The term 'Optimistic' is based on this assumption.

Since previous FTL algorithms did not consider the ex- The Optimistic FTL is much simpler than the previousistence of the NVRAM write buffer, their design policy log block based FTL algorithms. It has only one log block'efficiently accommodate different write patterns (ex. ran- and maintains only block mapping table in the main mem-dom small writes or sequential large writes) to the flash ory. While the BAST and FAST algorithms maintain notmemory' added large complexity to the FTL. For example, only a block mapping table but also a sector mapping tableBAST [1] and FAST [2] algorithms use log blocks to cope for the log block, Optimistic FTL does not maintain a sec-with random small writes. A page is not updated but is in- tor mapping table since it does not use sector mapping forvalidated and the new version of the page is written in the the log block. Also, instead of three types of merge opera-log block. tions - switch, smart copy and switch copy, Optimistic FTL

Using an NVRAM write buffer for page clustering has only one type of merge operation: Instant merge. Fig-(where pages in the buffer are clustered according to the ure 1 shows the instant merge operation. Assume that thereblock number in the flash memory and a page cluster is se- are eight pages in a block and currently all pages in a datalected as a victim for replacement), small random writes block are filled with valid data pages. The write requestcan be transformed into sequential writes to the flash mem- queue in the figure shows the sequence ofvictim pages thatory [3]. If we can delay writing a page cluster until the should be written to the flash memory. We can see thatcluster become sufficiently large, it is not necessary to de- pages 3, 4 and 7, which belong to the same data block inlay erasing data blocks in the flash memory because many the flash memory, are selected together as victim pages, inof the pages in the data block are already obsolete and the order of page number, according to the block based writenumber of valid pages in the original data block, which buffer management policy. When a write request to pageshould be copied to the new data block, is small. If we 3 arrives, Optimistic FTL copies pages 1 and 2 from data

*This work was supported by grant No. RO1-2007-OOO-20649-O from block to the log block ( I, O) and then writes page 3 inthe Basic Research Program of the KOSEF. the buffer to the log block((c)). On receiving a write re-

tCorresponding author quest to page 4, it writes the page into the log block at

Page 2: [IEEE Telecommunication Systems (MASCOTS) - Baltimore, MD, USA (2008.09.8-2008.09.10)] 2008 IEEE International Symposium on Modeling, Analysis and Simulation of Computers and Telecommunication

Data block Log block 1.4

....1..2 Valid page copy a Erase

2 0 vi~~~~~~~~~~~~ctim cluster30~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~............~~~~~~~~~~~~~~~~..8

02 u r n ... E u mE~~~~~~~~~~~~~~~~~~~~~...... .... ... .u...0 -~~~~~~~~~~~~~~~~~~~~~.. .. .. ... ...... ......

Fre 2 M0.. 1162612 116326 12.163642of. log blocks..

.. MB.1 B.2M.....A Size

Figure~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~)04_1.Instan.merge.oeration. alid.pagcopy ..Erasonce since all preceding pages are a~~~~lready--written..to.thelog block (®). Whenawrite request to page 7 arrives, it~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~... .......copies pages 5 and 6 from the data block to the log block~~~~~~~...0.3~~~~~~~~~~~~~~~~~~.. ..~~ ~~ and then writespage7 in the buffer to the log block o.~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ~~* * *... NKUUUliii*~~~~~~~~~~~~~~~~~~.......(0). Whenawriterequesttopage 12 arrives, since page ~~~~~~~~~~~~~~~~~~~~~ 111111111 111111111 liii iiiii~~~~..l2belongstoadifferentdatablockthanthedatablockfor~~~~~~~~~~~~~0. ..the previously requestedpages, Optimistic FTL copies the Z E E E~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~.......

table merits compared to the previous log block based FTL~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~co 'mU_

stablemerge lat~~~~Eray. agedsytemmno onyneestatsawoebfe-wr TFree block Data blolgoithcfrkprfomaceimrv enbualocns-

3. Performance Evaluation plify the FTL algorithm. In addition, Optimistic FTL can

tmsiFT.Figure 2.Ishowsth perfeoprmanceof three log;9p_ Eas

tim siefo vald pagecopiespandserae opreratinsw hie t hae oprto, u lo rvdslo0n tal egeltny

thegoverheadhofOptimisticeFTL swhen wriebuffersizes ist eerne0.5MB is 1. Aswe can see from the figure, Optimistic~~~~, 0. ... ...

FTL outperformsBAST in all cases and FAST when the [1] J. Kim, J. M. Kim, S. H. Noh, S...L..M.,.and...Cho.. ANVRAM size is lrger than or eual to 8 Mbyte, regardless sace-efficient lash translatin layer for.copact.flash sys

comispared to FaST,rotheOptmitic FTloc noth lonl shows g .y....Embdde Computing System, .6(3) 2007.very stable latency but also the maintenance overhead is [3] 5. Park, H. Jung, H.3.Shim,.S..Kang,and.J.. Cha..Using.non-little. We can conclude, based on the results, that using volatile ram as a write buffer for nand flash.. memory-basedNVRAM as a write buffer for a flash memory-based stor- storage devices. In Proc... of.MASCOTS..,.....Sept.. 2008..