WAFTL: A Workload Adaptive Flash Translation Layer with Data Partition

Transcription

1 WAFTL: A Workload Adaptive Flash Translation Layer with Data Partition Qingsong Wei Bozhao Gong, Suraj Pathak, Bharadwaj Veeravalli, Lingfang Zeng and Kanzo Okada Data Storage Institute, A-STAR, Singapore National University of Singapore MSST 2 May 26, 2

2 Introduction Outline Flash Fills the Performance Gap IOPS,,, s Processor DRAM memory Much faster than Disk(X-X) Non-volatile (Disk like) Low power consumption Light weight Resistance to shock,, s, s s Flash memory Hard Disk Picoseconds Nanoseconds Microseconds Milliseconds Latency

3 Introduction Outline NAND Flash Memory Different access unit Read/write in pages (us) Erase in blocks (very slow, ms) Out-place-Update: Does not allow overwrite. Limited number of erase per cell. -3K for SLC and -3K for MLC. Poor random write performance. Page Block Block Page Block n- m=64/28/256 Page m- Flash Memory

4 Low-level Flash Controller Introduction Outline Flash Translation Layer (FTL) in SSD Embedded Processor SRAM Control Bus NAND Flash NAND Flash NAND Flash LBA PBA Host Host Interface Data Bus Data Buffer Data Bus Flash Bus SATA, SCSI, FC, PCI etc Host I/F Layer Flash Translation Layer NAND I/F Layer NAND Flash NAND Flash NAND Flash

5 State-of-the-Art Outline Page-based FTL + High performance + High space usage + Low garbage collection overhead - Large mapping table - Not scalable Address Mapping table Flash LPN PPN Host Logical Page No (LPN) Block Block Physical Page No (PPN) LPN: Block 2 Block 3

6 State-of-the-Art Outline Block-based FTL + Small mapping table - Low performance - Low space usage (internal fragmentation) - High garbage collection overhead Host Logical Page No (LPN) Logical Block NO (LBN) Physical Block No (PBN) LPN: 7 LBN: 7/2=3 Offset: Address Mapping table LBN PBN PBN: Offset: Flash Block Block Block 2 Block 3

7 State-of-the-Art Outline Log-buffer based Hybrid FTL + Small mapping table + Improved performance comparing to Block-based FTL - Poor performance for random writes because of expensive merge operations Association One to one (BAST) One to Many (FAST, LAST) Many to Many (KAST) Data Block B B B2 B3 B4 L L Log Block Block-level mapping Page-level mapping Merge Operations

8 State-of-the-Art Outline : Demand-based FTL + Partial buffered page-based FTL + Good Performance (but lower than pure page-based FTL) + Reduced memory requirement - Large mapping table Page-based Mapping

9 State-of-the-Art Outline Page-based FTL Block-based FTL Hybrid FTL Performance High Good Low Middle Garbage Collection Overhead Low Low High Middle Favored Workload Random Random Sequential Sequential Mapping table Size Large Large Small Small Workload adaptive No No No No Scalable with SSD capacity No No Yes Yes Generally limited in one way or another, either in memory requirement, performance, garbage collection overhead or scalability.

10 WAFTL: Outline Workload Adaptive FTL Design Rationale Data Data Block-level Mapping Blocks (BMB) Page-level Mapping Blocks (PMB) Block... Flash Memory

11 WAFTL: Outline Workload Adaptive FTL Design Rationale Buffer Zone Page-level Mapping Data Partition Data Data Block-level Mapping Blocks (BMB) Page-level Mapping Blocks (PMB)

12 WAFTL: Outline Workload Adaptive FTL Two-level Address Translation Process Page Mapping Table (PMT) LPN Global Mapping Table (GMT) LPN PPN 3 PPN... LBN PBN LBN =LPN/Blk_Size - Page Mapping Blocks Block Offset: LPN mod Blk_Size 3 2 PBN +... LPN: Logical Page Number PPN: Physical Page Number LBN: Logical Block Number PBN: Physical Block Number Block Offset Block Mapping Blocks

13 WAFTL: Outline Workload Adaptive FTL Layout and In-memory Data Structure Request SRAM Buffer Zone Mapping Tree (BZMT) Global Mapping Table (GMT) LBN PBN 2 - Cached Page Mapping Table (CPMT) LPN PPN Data access Mapping replacement M PPN = M PPN =5 LBN PBN LPN PPN Data M VPN = M VPN = BMB PMB GMT PMT Buffer Zone Data Blocks Mapping Table Blocks

14 WAFTL: Outline Workload Adaptive FTL Buffer Zone Mapping Tree Key: Block Number Root node Interior node Leaf node Page count Block Popularity Block Dirty Pointer Array Block Node LPN 2 22 PPN Dirty Page-level mapping Buffer Zone Page 4 of 9

15 WAFTL: Outline Workload Adaptive FTL Buffer Zone Migration Block No: Block No: 2 Block No: Popularity: 3 Page count:4 8 9 Popularity: 2 Page count: Popularity: 2 Page count: Most Popular Block Yes No Page count >=Threshold Yes No Buffer Zone Buffer Dirty Buffer Zone Zone Data Active Page Active BMB Active PMB BMB PMB

16 Mapping Table Size(MB) Average Reesponse Time(ms) WAFTL: Outline Workload Adaptive FTL Determining Migration Threshold Performance oriented threshold Threshold is statically set as full block size and only full block data is stored into BMB. BEST performance, but Big Mapping table Dynamic threshold considering memory constraint Initially, THR is set to Block size THR decreases to a smaller value if performance degradation is smaller than D, and mapping table reduction is larger than D. D is a control parameter defined by users Fin Fin Threshold(pages) P P Blk _ size M M Blk _ size P THR M P P Blk _ size Blk _ size M Blk _ size M Blk _ size THR P THR and P Blk_size represent performance when threshold is set as THR and full block, respectively. M THR and M Blk_size represent mapping table size when threshold is set as THR and full block, respectively D D Threshold(pages) Fin Fin 2 Mapping table size is more sensitive to threshold than performance.

17 WAFTL: Outline Workload Adaptive FTL Garbage Collection Policy Requests Idle time Service time Offline Garbage Collection block erasure 2 blocks erasure block erasure Idle period (t predicti ) is predicted as t i predict t i real ( ) t i predict WAFTL use predicted idle length to calculate and determine how many invalid blocks (N i ) to be erased according to following Equation N i i t T predict erasure if if t t i predict i predict T T erasure erasure

18 Evaluation Outline Setup SSD Simulator FTL schemes: WAFTL, Pure Pagebased,, FAST 7 enterprise workloads Evaluation Metrics Average response time Erase count Page read/write operations Workload Traces Configuration Page Read to Register Page Program (Write) from Register Block Erase Serial Access to Register (Data bus) SSD Die Size Block Size Page Size Data Register Erase Cycles Buffer Zone sizes : 256MB/52MB/GB Workloads Avg. Req. Size(KB) Write(%) Seq.(%) Avg. Req. Inter-arrive Time(ms) Fin %.6% 8.89 Fin %.7%.8 Exchange %.5%.348 DevDiv %.9% DevDiv %.5% Sequential Read % 6% Sequential Write % 7% μs 2μs.5ms μs 2 GB 256 KB 4 KB 4 KB K SRAM for and WAFTL: 256KB For pure page-level FTL and FAST, we assume SRAM is enough to hold entire mapping table.

19 Average Response Time(ms) Erase Count() Page operations(million) Average Response Time(ms) Erase Count() Page operations(million) Evaluation Outline Pure Page FTL FAST WAFTL Zone52MB 32GB 48GB 64GB 8GB SSD Capacity WAFTL Zone=256MB WAFTL Zone=GB Pure Page FTL FAST WAFTL Zone=52MB 32GB 48GB 64GB 8GB SSD Capacity WAFTL Zone=256MB WAFTL Zone=GB Pure Page FTL Page write Page read Fin Trace (Static Migration threshold=64pages,256kb) FAST WAFTL Zone=GB.6.5 Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB 25 2 Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB 3 25 Page write Page read GB 48GB 64GB 8GB SSD Capacity 32GB 48GB 64GB 8GB SSD Capacity Pure Page FTL Fin2 Trace (Static Migration threshold=64pages,256kb) FAST WAFTL Zone=GB

20 Average Response Time(ms) Erase Count() Page operations(million) Average Response Time(ms) Erase Count() Page operations(million) Evaluation Outline Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB Pure Page FTL FAST WAFTL Zone=52MB WATL Zone=256MB WAFTL Zone=GB Page write Page read GB 48GB 64GB 8GB SSD Capacity 32GB 48GB 64GB 8GB SSD Capacity Pure Page FTL FAST WAFTL Zone=GB Exchange Trace (Static Migration threshold=64pages,256kb) Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB 2 Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB 2 Page write Page read GB 48GB 64GB 8GB SSD Capacity 32GB 48GB 64GB 8GB SSD Capacity Pure Page FTL FAST WAFTL Zone=GB DevDiv Trace (Static Migration threshold=64pages,256kb)

21 Average Response Time(ms) Erase Count() Page operations(million) Average Response Time(ms) Erase Count() Page operations(million) Evaluation Outline Pure Page FTL FAST WAFTL Zone=52MB 32GB 48GB 64GB 8GB SSD Capacity WAFTL Zone=256MB WAFTL Zone=GB Pure Page FTL FAST WAFTL Zone=52MB 32GB 48GB 64GB 8GB SSD Capacity WAFTL Zone=256MB WAFTL Zone=GB Pure Page FTL Page write Page read DevDiv9 Trace (Static Migration threshold=64pages,256kb) FAST WAFTL Zone=GB 6 5 Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB 3 25 Pure Page FTL FAST WAFTL Zone=52MB WAFTL Zone=256MB WAFTL Zone=GB 3 25 Page write Page read GB 48GB 64GB 8GB SSD Capacity GB 48GB 64GB 8GB SSD Capacity 5 5 Pure Page FTL Seq_Read Trace (Static Migration threshold=64pages,256kb) FAST WAFTL Zone=GB

22 Mapping table size(mb) PMB Blocks Distribution(%) Evaluation Outline Workload adaptive 8 256MB Buffer Zone 52MB Buffer Zone GB Buffer Zone WAFTL is adaptive to workloads (32GB SSD) Fin Fin 2 Exchange DevDiv DevDiv9 Seq-Read Seq-Write Traces 4 Pure Page Mapping FAST WAFTL Mapping table sizes under different workloads (256MB buffer zone, 32GB SSD) 3 2 Fin Fin 2 Exchange DevDiv DevDiv9 Seq-Read Seq-Write Traces

23 Mapping Table Size(MB) Average Reesponse Time(ms) Evaluation Outline Effect of Threshold Dynamic Threshold D=.6 Dynamic Threshold D=.3 Static Threshold THR= Buffer Zone: 256MB SSD: 32GB, 64Pages/Block Fin Fin 2 Exchange DevDiv DevDiv9 Seq-Read Seq-Write Traces 3 Dynamic Threshold D=.6 Dynamic Threshold D=.3 Static Threshold THR=64 2 4% 6% Buffer Zone:256MB SSD: 32GB, 64Pages/Block Fin Fin 2 Exchange DevDiv DevDiv9 Seq-Read Seq-Write Traces

24 Conclusion Outline WAFTL explores either page-level or block-level address mapping for data blocks based on access patterns in order to balance the mapping efficiency of a block-based approach and the GC efficiency of a page-based approach. a small part of NAND flash space for the buffer area to log data sequentially until it is full and then migrate them into either a page mapping block or block mapping block, depending on the migration threshold. different scheme to determine the value of migration threshold thereby, improving performance while reducing the mapping table size at the same time.

25