Flash SSD Operations

Sam Li

2022-04-28

Nand Flash SSDs¶

Terminology in flash: banks consist of blocks; and blocks consist of pages.
- banks/planes: consist of a large number of cells
- blocks: size 128 KB or 256 KB
- pages: size 4 KB
- cell level
  - one level: store a single bit into a transistor(1/0)
  - multi-level: store two bits into different levels of charge(00/01/10/11)… etc
- Flash-based SSDs
  - Flash translation layer(FTL)
  - Logical block address(LBA)
basic flash operations
- Read(a page)
- Erase(a block)
- Program(a page)
Page has state: INVALID, ERASED, VALID
raw flash performance
- device property: a page can be written until the entire block first be erased.
- Given above property: program-erase cycle
  1. latency: we should move data that is about to be erased to another place.
  2. Problem of reliability: wear out
FTL organization
- direct mapped: an operation to logical page N is mapped to physical page N.
  - Bad: causing the flash constantly overwritten and worn out eventually
- log structured: in-memory mapping table
  - Good:use logging to speedup recovery

Performance problem:

Flash operation: erase >> program(10*) >> read (10*)
- How to improve:
  1. parallel
  2. reduce write amplification
garbage collection(Log-structured FTL): old versions of data takes over the space.

Overwritten -> garbage data -> data migration
high cost of in-memory mapping table(Log-structured FTL)

Riliability problem:

Wear out

(Head crash in mechanical disk in contrast)
- wear leveling: spread writes across the blocks evenly
Disturbance: When accessing a page within a flash, some bits can get flipped in neighboring pages.
- sequential programming: write pages in order can reduce the disturbance.