SIAS-Chains: Snapshot Isolation Append Storage Chains Dr. Robert - PowerPoint PPT Presentation

SIAS-Chains: Snapshot Isolation Append Storage Chains Dr. Robert Gottstein Prof. Ilia Petrov M.Sc. Sergej Hardock Prof. Alejandro Buchmann

Motivation: Storage Technology Evolution Significant impact of storage technology evolution 30000 260 Sequential Throughput [MB/s] Random Throughput [IOPS] 240 read 220 write 200 180 3000 160 140 120 100 300 80 60 40 20 30 0 4 8 16 32 64 128 256 8KB 16KB 32KB 64KB 128KB 256KB 512KB 1024KB Blocksize [KB] Blocksize [KB] ▪ Intel X25-E SLC SSD ▪ Savvio 15k HDD ▪ Seq. Read / Write: 160 MB/s ▪ Seq. Read/Write: 250 / 170 MB/s ▪ Read/Write IOPS: 350 / 300 ▪ Read/Write IOPS (4K): 35 000 / 3 300 ▪ Latency Read/Write: 3.2 / 3.5 ms ▪ Latency Read/Write (4K): 0.075/0.085 ms ▪ Direct overwrite ▪ Erase before overwrite ▪ slow & large granularity 01.09.2017 | Dr. Robert Gottstein |

Motivation: Storage Technology Evolution Significant impact of storage technology evolution 30000 260 Sequential Throughput [MB/s] Random Throughput [IOPS] 240 read 220 write 200 HDD : symmetric read/write ; high Latency ; big block; 180 3000 160 rotational moving parts 140 120 100 300 80 SSD : asymmetric read/write ; low Latency ; No In-Place Updates ; 60 40 small block; write sequentialization; Intrinsic Parallelism ; Endurance 20 30 0 4 8 16 32 64 128 256 8KB 16KB 32KB 64KB 128KB 256KB 512KB 1024KB Blocksize [KB] DBMS needs to Leverage: Blocksize [KB] ▪ Fast Reads ▪ Intel X25-E SLC SSD ▪ Savvio 15k HDD ▪ Low Latencies ▪ Seq. Read / Write: 160 MB/s ▪ Seq. Read/Write: 250 / 170 MB/s ▪ Asymmetry ▪ Read/Write IOPS: 350 / 300 ▪ Read/Write IOPS (4K): ▪ Parallelism 35 000 / 3 300 ▪ Write Sequentialization ▪ Latency Read/Write: 3.2 / 3.5 ms ▪ Latency Read/Write (4K): 0.075/0.085 ms ▪ Direct overwrite ▪ Erase before overwrite ▪ slow & large granularity 01.09.2017 | Dr. Robert Gottstein |

Motivation: Storage Technology Evolution Significant impact of storage technology evolution 30000 260 Sequential Throughput [MB/s] Random Throughput [IOPS] 240 read 220 write 200 HDD : symmetric read/write ; high Latency ; big block; 180 3000 160 rotational moving parts 140 120 100 300 80 SSD : asymmetric read/write ; low Latency ; No In-Place Updates ; 60 40 small block; write sequentialization; Intrinsic Parallelism ; Endurance 20 30 0 4 8 16 32 64 128 256 8KB 16KB 32KB 64KB 128KB 256KB 512KB 1024KB Blocksize [KB] DBMS needs to Leverage: Blocksize [KB] ▪ Fast Reads ▪ Intel X25-E SLC SSD ▪ Savvio 15k HDD Multi Version DBMS: In principle suitable for asymmetric storage. ▪ Low Latencies ▪ Seq. Read / Write: 160 MB/s ▪ Seq. Read/Write: 250 / 170 MB/s ▪ Asymmetry Parallelism. Out-of place updates. Sequentialization.... ▪ Read/Write IOPS: 350 / 300 ▪ Read/Write IOPS (4K): ▪ Parallelism 35 000 / 3 300 ▪ Write Sequentialization ▪ Latency Read/Write: 3.2 / 3.5 ms ▪ Latency Read/Write (4K): 0.075/0.085 ms ▪ Direct overwrite ▪ Erase before overwrite ▪ slow & large granularity 01.09.2017 | Dr. Robert Gottstein |

Introduction Version Organization & Invalidation Relation R … A …. W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Version X 0 … … Item X 9 Version X 1 … … 10 Tuple X 0 Value=9 Tuple X 0 Value=9 Version X 2 … … 11 (ts create =123, ts inval =134) (ts create =123, ts inval =null) Visibility Tuple X 1 Value=10 Tuple X 1 Value=10 ▪ Timestamps (ts create =134, ts inval =null) (ts create =134, ts inval =141) creation: ts create invalidation: ts inval Tuple X 2 Value=11 (ts create =141, ts inval =null) ▪ Asymmetric : Fast Reads & Slow Writes ▪ Low Latency : no moving parts ▪ No In-Place Updates : Need to erase first (slow) ▪ Intrinsic Parallelism : Read in parallel 01.09.2017 | Dr. Robert Gottstein |

Introduction Version Organization & Invalidation Relation R … A …. W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Version X 0 … … Item X 9 Version X 1 … … 10 Tuple X 0 Value=9 Tuple X 0 Value=9 Version X 2 … … 11 (ts create =123, ts inval =134) (ts create =123, ts inval =null) Visibility Tuple X 1 Value=10 Tuple X 1 Value=10 ▪ Timestamps (ts create =134, ts inval =141) (ts create =134, ts inval =null) creation: ts create invalidation: ts inval Tuple X 2 Value=11 (ts create =141, ts inval =null) ▪ Asymmetric : Fast Reads & Slow Writes ▪ Low Latency : no moving parts Version Organization & Invalidation ▪ No In-Place Updates : Need to erase first (slow) Small Random Updates ▪ Intrinsic Parallelism : Read in parallel 01.09.2017 | Dr. Robert Gottstein |

SIAS: Snapshot Isolation Append Storage SIAS in a nutshell: redesign architecture and algorithms Tuple X 0 Value=9 W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Tuple X 0 Value=9 (ts create =123, ts inval =null) (ts create =123, ts inval =134) ▪ Version Organization Tuple X 1 Value=10 Tuple X 1 Value=10 (ts create =134, ts inval =null) (ts create =134, ts inval =141) ▪ Backward Chaining of versions ▪ Chain identified by virtual ID ( VID ) Tuple X 2 Value=11 ▪ Store the entrypoint in datastructure: VID map (ts create =141, ts inval =null) ▪ New Invalidation Tuple X 2 Value=11 ▪ Invalidation coded within the chain (ts create =141, VID=34) ▪ „One - place“ Invalidation Tuple X 1 Value=10 (ts create =134, VID=34) ▪ Append Storage Tuple X 0 Value=9 Item X ▪ Append tuple versions to a new page (ts create =123, VID=34) VID=34 ▪ Write page when filled or on a threshold 01.09.2017 | Dr. Robert Gottstein |

SIAS: Snapshot Isolation Append Storage SIAS in a nutshell: redesign architecture and algorithms Tuple X 0 Value=9 W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Tuple X 0 Value=9 (ts create =123, ts inval =null) (ts create =123, ts inval =134) ▪ Version Organization Tuple X 1 Value=10 Tuple X 1 Value=10 (ts create =134, ts inval =null) (ts create =134, ts inval =141) ▪ Backward Chaining of versions ▪ Chain identified by virtual ID ( VID ) Tuple X 2 Value=11 ▪ Store the entrypoint in datastructure: VID map (ts create =141, ts inval =null) ▪ New Invalidation Tuple X 2 Value=11 ▪ Invalidation coded within the chain (ts create =141, VID=34) ▪ „One - place“ Invalidation Tuple X 1 Value=10 (ts create =134, VID=34) Variant above is widely spread in multi version databases! ▪ Append Storage Variant below allows to address Flash storage properties Tuple X 0 Value=9 Item X ▪ Append tuple versions to a new page (ts create =123, VID=34) VID=34 ▪ Write page when filled or on a threshold 01.09.2017 | Dr. Robert Gottstein |

Multi Version DBMS Example W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Invalidation Creation Item X X 0 =9 X 1 =10 X 2 =11 Ti Transaction T1 T2 T3 DB Page P n P 0 P 10 P 21 P 32 P 4 P... Device Block B n B 0 B 10 B 21 B 32 B 4 B... 01.09.2017 | Dr. Robert Gottstein |

Multi Version DBMS Example W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Invalidation Creation Item X X 0 =9 X 1 =10 X 2 =11 Ti Transaction T1 T2 T3 DB Page P n P 0 P 10 P 21 P 32 P 4 P... Device Block B n B 0 B 10 B 21 B 32 B 4 B... 01.09.2017 | Dr. Robert Gottstein |

Multi Version DBMS Example W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Invalidation Creation Item X X 0 =9 X 1 =10 X 2 =11 Ti Transaction ▪ Random Writes T1 T2 T3 ▪ In-Place Updates DB Page P n ▪ Mixed Load P 0 P 10 P 21 P 32 P 4 P... Device Block B n B 0 B 10 B 21 B 32 B 4 B... 01.09.2017 | Dr. Robert Gottstein |

SIAS Principle Example Tuple Append Storage Management W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Invalidation Creation Item X X 0 =9 X 1 =10 X 2 =11 Ti Transaction VID Map T1 T2 T3 DB Page P n … X 2 X 1 X 0 P n Device Block B n … B k-3 B k-2 B k-1 B k … Write Order ▪ No in-place invalidation ▪ DBMS specific ▪ Append versions instead of pages ▪ Write reduction ▪ Write filled pages ▪ Simplyfied Buffer Management 01.09.2017 | Dr. Robert Gottstein |

SIAS Principle Example Tuple Append Storage Management W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Invalidation Creation Item X X 0 =9 X 1 =10 X 2 =11 Ti Transaction VID Map T1 T2 T3 DB Page P n … X 2 X 1 X 0 P n Device Block B n … B k-3 B k-2 B k-1 B k … Write Order ▪ No in-place invalidation ▪ DBMS specific ▪ Append versions instead of pages ▪ Write reduction ▪ Write filled pages ▪ Simplyfied Buffer Management 01.09.2017 | Dr. Robert Gottstein |

SIAS Principle Example Tuple Append Storage Management W 1 [X 0 =9];C 1 ; W 2 [X 1 =10];C 2 ; W 3 [X 2 =11];C 3 ; Invalidation Creation Item X X 0 =9 X 1 =10 X 2 =11 Ti Transaction VID Map T1 T2 T3 DB Page P n … X 1 X 0 X 2 P n Device Block B n … B k-3 B k-2 B k-1 B k … Write Order ▪ No in-place invalidation ▪ DBMS specific Significant Write Reduction ▪ Append versions instead of pages ▪ Write reduction (5 pages vs. 1 page) ▪ Write filled pages ▪ Simplyfied Buffer Management Sequentialization 01.09.2017 | Dr. Robert Gottstein |