Hig igh-Performance Key- Carnegie Mellon Value Store University - - PowerPoint PPT Presentation

SIL ILT: A Memory ry-Efficient, Hig igh-Performance Key- Value Store

Hyeontaek Lim, Bin Fan, David G. Andersen Michael Kaminsky† Carnegie Mellon University †Intel Labs

Presented by: Ramya Danappa

Seesaw Game?

Memory efficiency High performance

2

FAWN-DS HashCache BufferHash FlashStore SkimpyStash How can we improve?

2 4 6 2 4 6 8 10 12

(1) “The memory overhead and lookup performance of SILT and the recent key-value stores. For both axes, smaller is better.” Exp Explain the the po positions of

• f FAWN-DS,

S, Ski SkimpyStash, , Bu BufferHash, and and SIL SILT T on

• n the

the grap aph.

Read amplification Memory overhead (bytes/entry) FAWN-DS HashCache BufferHash FlashStore SkimpyStash

3

SILT

Describe SILT’s structure (Architecture of SILT). Compared with LevelD lDB, SIL ILT T has has onl

• nly thr

three levels. What’s concern with a multi-le level KV KV stor

• re whe

hen it it ha has too

• o few le

levels?

“SILT uses a memory ry-effi ficient, hig igh-performance hash sh table le base sed upon cuckoo hashing.”. Explain what the cuckoo hash shin ing is is and why y it it is is use sed.

• R. Pagh and F. Rodler.

Cuckoo hashing. Journal

• f Algorithms, (2): 122–

144, May 2004.

Desig ign of f LogStore: an in in-memory cu cuckoo hash table (in (index and fil filter) to describe how a PUT request and a GET request is is served in in a

• LogStore. In

In particular, explain how th the tag is is used in in a LogStore for cu cuckoo hashing

Explain how a LogStore is converted into a HashStore?

“Once a LogStore fills lls up up (e (e.g., the insertio ion alg lgorithm terminates without finding any vacant slo lot after a maxim imum number of

• f dis

ispla lacements ts in in the hash table), SILT freezes the LogStore and converts ts it it into a more memory ry-efficie ient data str tructure.” Compared to to LogStore, what’s th the advantage of

• f HashStore? Why doesn’t SILT create HashStore at

at the beg beginnin ing (w (wit ithout fi first cr creating Log LogStore)? )? Advantage of Hashstore over logstore:

• Hashstore saves memory over Logstore by eliminating the index and reordering the on-flash

(key,value) pairs from insertion to hash order. SILT create Logstore first because so every data ideally should store on flash.

“When fixed-le length key-valu lue en entrie ies ar are sort

• rted by

y key on

• n fl

flash ash, a a tri trie for

• r the

the sho hortest uni unique pr prefix fixes of

• f the

the keys ser erves as as an an index for these sorted data.” While a So Sort rtedStor

• re is ful

fully ly sort

• rted, cou
• uld

ld you

• u comment on
• n the

the cos

• st of
• f me

mergi ging g a a Has HashStor

• re wi

with th a a So Sort rtedStor

• re? Co

Comp mpare thi this cos

• st to
• the

the ma majo jor com

• mpaction
• n cos
• st for
• r Le

LevelDB?

• Cost is proportional to the size of Database.
• SILT is worse because single sorted store leads to bad write amplification.

Questions?

Thank you