연세대학교 Yonsei University Data Processing Systems for Solid State Drive Yonsei University Mincheol Shin

Overview Main Target : Data Processing Systems with SSD Purpose : Improving I/O Performance Data Processing System – Relational Database Management System e.g. Oracle, MySQL, PostgreSQL, SQLite – Distributed Data Processing System e.g. Hadoop Distributed File System, MapReduce, Hive, Hbase, Tajo, Spark – Key-value Store e.g. Redis

Outline Solid State Drive (SSD) RDBMS on Solid State Drive Big Data Processing for Solid State Drive

Solid State Drive: Flash Memory [VLDB2011Tut2] Great Performance !! – High I/O Performance: 41 MB/s Read, 7.5 MB/s Program [Micron 2014] – Fast Random Access: Under 0.1 ms (HDD: 2.9 to 12 ms) – Low Energy Consumption Four Constraints of NAND Flash Memory – C1: Program granularity (2KB~16KB) – C2: Must erase a block before updating a page (256KB ~ 1MB) – C3: Pages must be programmed sequentially within a block – C4: Limited lifetime (10 4 ~ 10 5 ) 4k Page A Erase Block (1 MB) [VLDB2011Tut2] P. Bonnet, L. Bouganim, I. Koltsidas, S. D. Viglas, VLDB 2011 Tutorial: System Co-Design and Data management for Flash Devices

Solid State Drive Solid State Drive (SSD) – Definition: Persistent data storage without disks nor a drive motor. – Support Traditional Block I/O Characteristics for SSD – Fast Random Access (inherited from flash memory) – Read/Write Imbalance (inherited from flash memory) – Exploiting Internal Parallelism (SSD internal structure) – In-Storage Processing SSD Host I/F (SATA, SAS, PCIE) Read(addr) Write(addr, data) Internal Algorithm (FTL) Map ping Wear leveling Garbage Collection Physical Storage Flash Chips Read Program Erase

Solid State Drive: Flash Translation Layer (FTL) Flash Translation Layer – Convert the block I/O operations to internal operations – Three Major Components Mapping – Map Logical Block Address(LBA) to physical page Garbage Collection Wear Leveling – To extend lifetime of SSD Logical Physical Block 1 Block 2Block 3Block 4 Update vvvvIIvIvv Block 2Block 3Block 4 vvvvIIIIvvv Block 2Block 3Block 4 Erase

Solid State Drive: Internal Parallelism SSD can read/write the data in parallel SSD Host I/F (SATA, SAS, PCIE) Flash Package Channel-level Parallelism (N Parallel Channels) Package-level parallelism (Interleaving) Memory Time Read 1 Transfer 1 Read 3 Transfer 3 Read 5 Transfer 5 Read 7 Transfer 7 Read 2 Transfer 2 Read 4 Transfer 4 Read 6 Transfer 6 Read 8 Transfer 8 Package 1 (Ch. 1) Package 2 (Ch. 1) Package 3 (Ch. 2) Package 4 (Ch. 2) Channel 1 Channel 2 Data 2Data 4Data 6Data 8 Data 1Data 3Data 5Data 7

Solid State Drive: Internal Parallelism Using internal parallelism, SSD achieves – High performance for sequential I/O Similar to Striping (RAID 0) Seq. bw for SATA SSD – Write : 450 MB/s – Read : 500 MB/s – High performance for concurrent I/O [VLDB2012Roh] H. Roh, S. Park, S. Kim, M. Shin, S-W. Lee, B+-tree index optimization by exploiting internal parallelism of flash-based Solid State Drives

Solid State Drive: In-Storage Processing SSD has CPU and Memory for FTL Host Interface is bottleneck ! – H/I has lower bandwidth than internal bandwidth of SSD Two approaches – Light-weight filter in SSD Transfer smaller data through H/F Filter tuples using predicates – Sub-modules in SSD e.g. Transaction management with COW Need special SSD to implement ISP – OpenSSD, SmartSSD and so or

DBMS on Solid State Drive Main research areas: – Buffer Management – Index Management – Query Processing – Transaction Management Most of researches using SSDs focused on storage I/O

DBMS on Solid State Drive: Index Management FD-tree – Exploit sequential bandwidths of SSDs – B-Tree + sorted runs PIO B-tree – Exploit internal parallelism of SSDs – Access to multiple B-tree node along multiple paths

DBMS on Solid State Drive: Query Processing FlashJoin: PAX based query processing – NSM layout Most typical page layout Tuples are stored in a contiguous region – PAX layout Values of columns are stored in contiguous region (minipage) Originally, PAX is designed for reducing cache miss in CPU cache – FlashScan reads only needed minipages – FlashJoin joins minipages read by flashScan

DBMS on Solid State Drive: Query Processing FMSort – Exploit internal parallelism of SSD – During merge phase,

DBMS on Solid State Drive: Transaction Mgmt. X-FTL: Shadow Paging in SSD – Writing operations of SSD is similar to Copy-on-write When a page is updated, the modified page is written to an empty page. And then, invalidate old page – X-FTL maintains old pages until transaction is committed. – There is no copying the original pages

Big Data on Solid State Drive 3 approaches to improve performance using SSDs – Complete replacement Higher cost per capacity – Selective replacement e.g. intermediate results on SSDs, HDFS data on HDDs – SSD as a cache Commercial/Noncommercial cache SW exist Open source : bcache, flashcache, enhanced IO, DM-cache Project with SK Telecom Archival Storage of HDFS – Store replica into 4 tiers of storage ARHIVE : slowest and biggest capacity storage (petabyte of storage) DISK, SSD, RAM_DISK hdfs/ArchivalStorage.html#Storage_Types:_ARCHIVE_DISK_SSD_and_RAM_DISK hdfs/ArchivalStorage.html#Storage_Types:_ARCHIVE_DISK_SSD_and_RAM_DISK Issues – Industry leads Big Data processing platform area – There is no standard model – Because CPU overhead are too high