Cost-effective Hybrid Storages

Cost-effective Hybrid Storages

Flash Group, Cao Qingling

Motivation

• Dry up things with the least money！

Motivation

• High cost, low density, low reliability.• Replacement as HDD is not recommended, especially

when the data volume is very large.• Cache as HDD, make up the gap between RAM and

HDD. Hurt the lifetime.• A permanent store at the same level as HDD, store

some special data.

Introduction

• Put forward a Hybrid Hbase with SSD.• Storing system component of Hbase in SSD,

which at the same level as HDD.• Perform quantitative assessment, Hybrid

Hbase perform 1.5-2 times better.

HBase

• Column-based key-value store.

• Each region server has a write-ahead log(WAL).

• First write WAL and then in-memory memstore.

• Region is a horizontal division.

• A region could split.

Data on disks is stored as Log-structured merge(LSM) trees.

HBase System Component

Zookeeper:• Clients contact it for -ROOT- table.• Master contacts it to know available region servers.• Region servers contact with it in a heartbeat keep-

alive mechanism.• Zookeeper is I/O intensive.Catalog Tables:• -ROOT- and .META. Tables.• Mostly read intensive and are not updated frequently.

HBase System Component

Write-ahead-log(WAL):• Any write is first done on the WAL.• The size grows with: i) WAL committed; ii) write rate;

iii) the size of key-value pair.Temporary Storage:• Used when a region is split or merged.• Sequentially read or write.

Assessment

1% of the database size. Gain more than 10% performance.

Price: 1:10

Experimental Evaluation

• Experiment: Intel processor(4 cores and 4 threads at 3 GHz) with 8 GB RAM, Western Digital 1TB HDD, Kingston 128 GB SSD.

• Yahoo! Cloud Serving Benching(YCSB).• Workloads: 100w queries on database with

6000w records. Record size is 1KB. Totally 72 regions.

Experimental Evaluation

Experimental Evaluation

Introduction

• Approximate membership query data structure(AMQ). Bloom Filter.

• Larger than RAM, performance decays.• Quotient Filter： better data locality,

squential operations, available delete, dynamically resized, space-saving.

• Buffered Quotient Filter(BQF) and Cascade Filter(CF) designed for flash.

Introduction

• Approximate membership query data structure(AMQ). Bloom Filter.

• Larger than RAM, performance decays.• Quotient Filter： better data locality,

squential operations, available delete, dynamically resized, space-saving.

• Buffered Quotient Filter(BQF) and Cascade Filter(CF) designed for flash.

Quotient Filter

• fr = f mod 2r

• fq =

• T[fq] = fr Fingerprint: f = fq2r + fr.

Quotient Filter

• is_occupied: check if fq = i, namely if T[i] has data.

• is_shifted: if fr belongs to slot i.• is_continuation: if blongs to the same run with i-1.

Physical Storage

run

Quotient Filter

• Check if f in the QA: step1: step2: to the beginning of the cluster. step3: to the start of the run. step4: search f. • Insert a f.• Delete a f.

Quotient Filters on Flash

• Buffered Quotient Filter - BQF: one QF as the buffer, another on SSD. - Optimized for lookup performance.• Cascade Filter - Optimized for insertion. - Offer a trade off between lookup and insertion.

Quotient Filter on Flash

• Cascade Filter - Based on cache-oblivious lookahead arrary(COLA).

Evaluation

Conclusions

• Bloom Filter has wide use in key-value storage.• Change the way of thinking.• Gain inspiration from traditional algorithms of

database.• Design corresponding hybrid system by

applications.

Bloom Filter

• Initial state:

• Insert: H(1), H(b).

• Can not expend, support no delete, poor data locality.

Back