Lets Talk About Storage & Recovery Methods for Non-Volatile - - PowerPoint PPT Presentation

Let’s Talk About Storage & Recovery Methods for Non-Volatile Memory OLTP Database Systems

Andy Pavlo + Joy Arulraj Carnegie Mellon University

Winter Winter 2013: 2013: Fir irst B t Blood lood

• Initial evaluation of

existing DBMSs on Intel NVM SDV

• Results published

ADMS@VLDB’14

MyS MySQL L vs.

• vs. H

H-Sto Store

90% Reads / 10% Writes 50% Reads / 50% Writes

Summer Summer 2014 2014: : Fir irst t Blood lood, P , Par art II t II

• Evaluate storage and

recovery methods for NVM.

• Preparing SIGMOD’14

submission.

DBMS DBMS Tes estbe tbed

• Custom lightweight DBMS.

–Uses NUMA & PMFS interfaces. –No volatile DRAM.

• Partition-based locking CC.
• Pluggable architecture:

–Supports different storage engines.

Engine ngine #1 #1 – In In-pla place ce U Upda pdate tes

• Apply change to tuples directly.

–VoltDB with ARIES. –Table storage + write-ahead log. –STX B+Tree

Engine ngine #2 #2 – Copy

• py-on
• n-Write U

ite Upda pdate tes

• Make new copy before updating:

–Shadow paging using LMDB Persistent B+Tree. –No logging. –Background garbage collection.

Engine ngine #3 #3 – Log

• g-bas

based U ed Updat ates es

• Changes only written to log.

–Based on LevelDB’s LSM. –No table storage. –Background level compaction.

Stor Storage E Engin ngines

Tabl able St Stor

• rag

age Loggi gging Ex Examp mple

In-Place Yes Yes VoltDB Copy-on-Write Yes No LMDB Log-based No Yes LevelDB

NVM O Optimiz ptimized E Engine ngines

• Refactored engines to be

“pointer-oriented”.

• Extended Intel’s libpmem

allocation library.

–Added arena-based allocation. –Significantly improved throughput.

Expe xperime imenta tal E l Evalua aluatio tion

• Yahoo! Cloud Serving Benchmark:

–2 million records (~2GB) –Two workload mixtures –Two skew settings –1 million transactions

Expe xperime imenta tal E l Evalua aluatio tion

• NVM Latency Configuration:

–2x DRAM (~200ns) –8x results not shown.

• 8 partitions on 8 cores.

Throughput

• ughput

90% Reads / 10% Writes 50% Reads / 50% Writes

4x 4x

NVM R Read eads/Wr Write ites

90% Reads / 10% Writes 50% Reads / 50% Writes

Stores Loads

Reco ecovery T Time ime

Zero Recovery

Dis iscu cussio ion

• NVM engines outperforms

“traditional” engines:

–Higher throughput –Reduced wear on device.

• In-place performs best overall.

N-STORE STORE

nstore.cs.cmu.edu

Fall all 2014 2014: : N-Sto Store

• First DBMS for NVM-only
• perating environment.
• OLTP/OLAP hybrid

–Column-store that supports fast in- place updates.

• Indexing + Many-Core

Andy Pavlo

• Col. Stan

Zdonik Mike Stonebraker Justin DeBrabant Joy Arulraj Subramanya Dulloor Rajesh Sankaran Jeff Parkhurst

@ANDY_PAVLO

END