Durable NAND flash memory management

词条积累

1.NAND flash memory

http://www.searchstorage.com.cn/whatis/word_6052.htm

http://baike.baidu.com/subview/64506/5136742.htm?fromtitle=NAND%E9%97%AA%E5%AD%98&fromid=8051173&type=syn

2.FTL

http://hi.baidu.com/alan_ding/item/7895367e324d5e3ed7a89c70

3.bit flips

http://www.crifan.com/order_nand_flash_bit-reversed_bit_flipped/

4.

Abstract：

1.本文提出了pcm-ftl，以耐用的NAND快闪记忆体管理方案为基础，重新设计以满足减少PCM写活动需要嵌入式系统。

2.基本的思想:要尽量避免储在PCM中的位翻转（即，映射表）

3.pcm-ftl采用二级映射机制,在I/O访问行为的特征上最大以最大限度地减少PCM上的写活动

PCM-FTL employs a two-level mapping mechanism, which focuses on minimizing write activities to PCM by characterizing I/O access behaviors

一、 INTRODUCTION

1.传统：

（1）功能增加，耗电增加

（2）DRAM时，断电后，重新遍历flash建立map-table

（3）DRAM的耗电和可扩展性到了极限

2.PCM

（1）PCM的主要问题，FTL写的频率很高——》减少不必要的写

（2）已经有的硬件优化：写前读

3.PCM-FTL

（1）管理主存的底层架构，写感知

PCM-FTL enhances the lifetime of PCM by making the management of NAND flash memory aware of write activities on underlying memory architecture.

（2）不改变文件系统和硬件

（3）PCM-FTL的基础思想是减少位翻转在PCM上

The basic idea is to try to avoid bit flips in the FTL mapping table, which is stored in PCM. By reducing write activities to PCM cells, the lifetime of PCM is enhanced

（4）页级映射：处理随机请求（此请求不频繁）

PCM-FTL uses page-level mappings to handle infrequent random requests

（5）块映射：处理顺序请求（此请求频繁）

a tiny amount of block-level mappings for most frequently accessed sequential requests

（6）PCM-FTL actively chooses a physical block in NAND flash memory whose physical block number incurs the minimum number of bit flips in the mapping entry when allocating a physical block

（7）With appropriate wear leveling algorithms, PCMFTL evenly distributes write activities across the whole PCM chip.

4.本文贡献

（1）present for the first time a write-activity-aware,two-level flash memory management technique

（2）use an effective wear leveling algorithm

（3）realistic I/O traces 实验

二、BACKGROUND AND MOTIVATION

1.hFTL

（1）a page-level mapping table in PCM keeps track of mappings between logical page number (LPN) and physical page number (PPN)

——》the mapping table is updated frequently and thus imposes the endurance issue for PCM

三、PCM-FTL

1.The basic idea of PCM-FTL is thus to preserve each bit in FTL mapping table,which is stored in PCM

2.A two-level mapping mechanism is proposed to separate the two types of request. For random requests, PCM-FTL uses a page-level mapping table (PMT)
to store their mappings. For large, sequential accesses, we use a small block-level mapping buffer (BMB) to cache the mappings.

3.With the consideration of write activities, once a block is needed for incoming write requests, PCM-FTL selectively allocates a physical block in NAND flash memory whose physical block number incurs minimum number of bit flips in PCM cells.

4.I/O workload=small random+ large sequential requests

5.Requests that intend to write more than t pages are categorized as sequential ones

6.

Random requests: PCM-FTL sequentially allocates physical pages from the first page of a physical block in NAND flash memory, so that all pages in blocks are fully utilized. Accordingly, PCM-FTL adds LPN to PPN mapping for random requests in PMT

Sequential requests: PCM-FTL allocates physical pages based on block offset as most sequential requests usually occupy a whole block, so that all pages in blocks are fully utilized as well. Similarly, PCM-FTL adds an LBN to PBN mapping in the BMB

7.Different from a traditional block-level scheme, the BMB in PCM-FTL is only for sequential requests and has limited capacity：The purpose is
to only buffer the mapping for the most frequently and recently updated blocks

8.handling writerequests

（1）New Requests：

If the request is a sequential one, pages will be written according to their offsets in the block and a block-level mapping is set up in the end. Random requests will be handled by writing to the newly allocated block one by one from the first physical page, regardless of the logical page number offsets. Page-level mappings will be set up.

（2）Update Requests：relies on the request type and the old block type.

连续请求直接分配一个新块，快号要和原来块映射里的保持最小翻转

随意请求，看原来的块有没有位置，一个接一个的写

连续请求，如果原来的块的空闲页足够的话，变成页映射

if the request is a sequential one, a new block will be allocated. In case the old block was mapped at block-level,
the allocation routine will try to find a block that will cause the minimum number of bit flips in the block table entry.

四、Algorithm

Algorithm 3.1 The wear leveling algorithm in PCM-FTL (pcm-ftl磨损均衡算法)

备注:

block-level mapping buffer (BMB)

page-level mapping table (PMT)

BMB records （记录）logical block to physical block mappings, while PMT records page-level mappings (i.e.,logical page to physical page)

对应论文3.4

(1) 在PMT中定期移动 BMB

疑问：

（1）PTE 、BTE

（2）Total PCM length

3.2 Handling new requests in PCM-FTL

Logical Page Number (LPN)

sequential 连续的

LBN/PBN: Logical/Physical Block Number

LPN/PPN: Logical/Physical Page Number

3.3 Handling update requests in PCM-FTL