Efficient Identification of Hot Data for Flash Memory Storage Systems

ACM Transaction on storage, Feb. 2006

JenWei Hsieh (National Taiwan University),LiPin Chang (National Chiao-Tung University),and TeiWei Kuo (National Taiwan University)

Speaker: Soyoon Lee

Hot data

• Frequently access data.• Usages

- garbage collection- performance of the flash memory access- longer lifetime ( wear leveling)

A Multihash-Function Framework- use independent multiple hash functions- reduce the chances of false identification of hot data- on-line- provide excellent performance- reduce the hash table space

Highly Efficient Method On-line Hot Data Identification!

non hot data hot data

• Garbage collection• Wear leveling• Physical/logical mapping

• lower level functionalities• read/ write/ erase

The count updating of an LBA

Hot data identifier

FTL

block write ( LBA , size ) from File

System

x

Logical Block Address x

f1(x)

f2(x)

f3(x)

f4(x)

K-dimensionalHashing Table

Hash Functions

H Most Significant Bits

+1

+1

+1

+1

y +1

The count updating of an LBA

• Freeze a counter once it reaches the maximum positive values

• Right-shift all counters for 1 bit whenever any counter is overflowed

• Exponential decaying

• When? : per decay period

• How? : the values of all counters divided by two Right-shift all counters

Count Overflow Decaying of count

The hot data identification of an LBA

Hot data identifier

FTLx

Logical Block Address x

f1(x)

f2(x)

f3(x)

f4(x)

K-dimensionalHashing Table

Hash Functions

H Most Significant Bits

block write ( LBA , size ) from File

System

Hot data H MSB of every counter of hashed values contain a

nonzero bit value

X is not hot data

Experimental setup• A Multihash-Function Framework setup

- The number of hash functions: 2 (the division method, the multiplication method)- Each counter for a hash table entry: 4 bits - hash table size: 1~8 KB

• Flash memory size: 512MB

• Traces of data access: the Hard disk of a mobile PC (a 20GB hard disk, 384MB RAM, Intel Pentium-III 800MHz)

• Comparison method- direct address method*M.L Chiang, C.H. Paul, and R.C. Chang, “Manage flash memory in personal communicate devices,” Procee

dings of International Symposium on Consumer Electronics,1997.

- two-level LRU*CHANG,L.P. AND KUO, T. W. “An adaptive striping architecture for flash memory storage systems of embed

ded systems. In Proceedings of the 8th IEEE Real-Time and Embedded Technology and Applications Symposium, 2002.

The locality in data access

• Direct address method• No false hot data identification• Decay period: 5117 writes• hot data threshold on the number of writes to an LBA: 4

Performance EvaluationHash table size vs False Identification

1 KB: basic

1 KB: enhanced

2 KB : basic

2KB enhanced,4KB basic,

8 KB enhanced

• basic: multi-hash function framework• enhanced: the framework with an enhanced counter update policy

512 flash memory, 2KB hash table 0.871% !!

Performance Evaluation

Decay period (setup: 5117 writes)

6396 writes

Performance Evaluation

Runtime Overheads

CPU Cycles per Operation (Unit: CPU Cycles)

The runtime overheads on the maintenance of the hash table (i.e., the decay process) is also limited

(e.g.,3,565 CPU cycles per 5,117 writes over 512MB flash memory).

Two – level LRU

Two-level LRU: hot data 512, candidate data 1024 nodes

Thank you!