Solid State Drive Feb 15 NAND Flash Memory

Solid State Drive Feb 15

NAND Flash Memory l Main storage component of Solid State Drive (SSD) l USB Drive, cell phone, touch pad…

Solid State Drive (nowadays)

Solid State Drive (SSD) architecture

Comparison

Comparison Attribute SSD HDD Random access time 0. 1 ms 5 -10 ms Bandwidth 100 -500 MB/s 100 MB/s sequential Price/GB 0. 9$-2$ 0. 1$ Size Up to 2 TB, 250 GB common 4 TB Power consumption 5 watts Up to 20 watts Read/write symmetry No Yes Noise No Yes (spin, rotate)

Characteristics of SSD l l High reliability (no moving parts) Small form factor Less noise Read/Write latency w. r. t. HDD l l Out-of-place update l l Erase at the unit of block LBN to PBN mapping (FTL) l l l Invalidate old version, and write new version somewhere else Invalidated version becomes garbage HDD, in-place update Garbage collection (GC) l l No seek latency Emulate a block device Erase cycle limit (endurance) Read latency 25 us Write latency 0. 25 ms Erase latency 2. 5 ms

l Advantages of NAND SSD l l l Widely deployed in high-end laptops l l Fast random read (25 us) Energy efficiency High reliability (no moving parts) Small form factor Less noise Macbook air, Think. Pad X series, touch pad… Increasingly deployed in enterprise environment either as a secondary cache or main storage

l Disadvantages of SSD l Garbage collection (GC) problem of SSD l l l Blocks in the SSD have a limited number of erase cycles l l l Stemmed from the out-of-place update characteristics Update requests invalidate old version of pages and then write new version of these pages to a new place Copy valid data to somewhere else (increasing number of IOs) Garbage collection is periodically started to erase victim blocks and copy valid pages to the free blocks (slow erase: 10 x. W, 100 x. R) 100, 000 for Single Level Chip (SLC), 5, 000 -10, 000 for Multiple Level Chip (MLC), can be as low as 3, 000 May be quickly worn out in enterprise environment Performance is very unpredictable l Due to unpredictable triggering of the time-consuming GC process

Drive read performance

Flash Translation Layer (FTL) l Page-mapping l Block-mapping l Log-block-mapping

Flash Translation Layer (FTL) l Page-mapping l l Mapping any Logical Block Address (LBA) to any Physical Page Address (LPA) Require huge RAM to store the mapping entries l 512 GB SSD requires 4 GB RAM to store the mapping table l Each mapping entry is 16 Byte, each page size is 2 KB

Flash Translation Layer (FTL) l Block-mapping l Map logical address to fixed offset of the physical address by taking module l LBN%N, N is the number of pages in each block l LBN with the same offset may conflict, high overhead

Flash Translation Layer (FTL) l Log-block-mapping

Issues about deployment Mixing read and write degrades the performance l Bandwidth drops as more data is being written l l l Garbage collection The performance becomes worse in enterprise environment Bandwidth can drop to 20% of the bandwidth of the datasheets Endurance l l Wear out in 23 days in enterprise environment MLC can only wear

Mixed writes and reads

Improve write performance l OS write caching l Flash specific file system l Drive Write Caching l Multiple concurrent erase blocks l Expensive: 200~400$ /GB

l Flash File System: ZFS, has be optimized l JFFS 2 l