Middleware and Flash Translation Layer Co-Design for the Performance Boost of Solid-State Drives

Transcription

1 Middleware and Flash Translation Layer Co-Design for the Performance Boost of Solid-State Drives Chao Sun 1, Asuka Arakawa 1, Ayumi Soga 1, Chihiro Matsui 1 and Ken Takeuchi 1 1 Chuo University Santa Clara, CA 1

2 Outline Introduction Three Techniques to Boost SSD Performance Storage Engine Assisted SSD (SEA-SSD) Logical Block Address (LBA) Scrambler Storage Class Memory (SCM)/ Flash Hybrid SSD Conclusion Santa Clara, CA 2

3 Solid-State Drive SSD is composed of flash memories. SSD controller manages the data storage. SSD Channel-1 Way-1 Way-2 Way-m Host interface logic SSD controller Way-1 flash Channel-2 flash Channel-n Way-2 flash flash Way-m flash flash Way-1 Way-2 Way-m DRAM flash flash flash Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 1, pp , Jan

4 Flash Memory The read/write unit of the flash memory is a page. The erase unit of the flash memory is a block. In-place overwrite is prohibited: overwrite involves a page read + a page write. Bit-Line (BL) Source-Line (SL) Page Block Word-Line (WL) Frequent overwrite create massive number of invalid pages. Santa Clara, CA [1] K. Takeuchi, JSSC, vol. 44, no. 4, pp ,

5 Garbage Collection of SSD GC is triggered when SSD runs out of free spaces. GC can be the bottleneck of SSD write throughput. Valid page Invalid page Blank page 1. Read valid pages in the next erase block 2. Data out/in, ECC Chip... Next erase block New block... Page Buffer 4. Erase the next erase block N valid : # of valid pages in the next erase block GC latency: N valid x ( T read + T write ) + T erase = N valid x 1.7ms + 8.5ms Page copy time Erase time If N valid exceeds 54, GC latency > 100ms 3. Write to a new block SSD Controller Santa Clara, CA [1] K. Takeuchi, JSSC, vol. 42, no. 1, pp ,

6 Outline Introduction Three Techniques to Boost SSD Performance Storage Engine Assisted SSD (SEA-SSD) Logical Block Address (LBA) Scrambler Storage Class Memory (SCM)/ Flash Hybrid SSD Conclusion Santa Clara, CA 6

7 Database Management system (DBMS) To solve the IO bottleneck of the DBMS system, high-speed and low-power storage is required. Solid-state drive (SSD) is a good fit. User application User application User application Database management system (DBMS) Database (in storage) Santa Clara, CA 7

8 Storage Engine Assisted SSD (SEA-SSD) OS is bypassed to reduce the communication overhead. Conventional scheme Proposed scheme App DBMS SE fopen () DBMS SE open_ssd() Hints OS FS GBD DD Logical address Bypassing OS layers Logical address SSD FTL Flash Physical address FTL Flash Physical address App: Application, DBMS: database management system, SE: Storage engine, FS: File system, GBD: Generic block device, DD: Device driver, FTL: Flash translation layer (SSD controller) Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 1, pp , Jan

9 SEA-SSD Architecture Three hints are passed from the storage engine (SE) to SSD controller. Data are classified by activity and stored to the dynamic and static segments of SSD, respectively. Hints (1) InnoDB SE (2) (3) Read/Write requests SE settings 0/1 Logical page addr. (2) Preliminary classification (1) Segment _dyn capacity estimation SSD controller SEG_dyn ( 1 ) Static data migration SEG_stat ( 0 ) Dynamic data prediction SSD Segment_dyn Segment _stat Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 1, pp , Jan (3)

10 SEA-SSD Management Algorithms Data activity is judged by Hint (2) and Hint (3) from SE. Dynamic data -> Segment_dyn; Static data -> Segment_stat. InnoDB SE Data from SE, Hints [Hint(2)+Hint(3)] Algorithm I Y Write to Segment_dyn Enough free pages? Y Hint (2)== SEG_dyn? N Write to Segment_stat N Y GC: garbage collection LPA: logical page addr. GC starts Copy page to Segment_dyn Valid page addr. in recycling block hits LPA hint [Hint(3)] list? GC done? Copy page to Segment_stat N N End Y Algorithm II Flash translation layer (SSD controller) Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 1, pp , Jan

11 25% 20% 15% 10% 5% 0% -5% -10% SSD Performance Evaluation Over 10% performance improvement is achieved for all workloads at 3 GB SSD capacity (33% flash over-provisioning). 30% SSD write performance gain <=50% Over-provisioning T1 T2 T3 S1 S2 S3 Realistic region for enterprise SSD Baseline = 0% SSD capacity (GB) Workload: TPC-C & Sysbench Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 1, pp , Jan

12 Short Summary SEA-SSD has been proposed to improve the write performance of the SSD for the database application. Hint information is passed from the SE to SSD controller for better classifying and predicting the data activity. A database application coupled simulation platform has been developed for such a cross-layer work, which accelerates the simulation speed by over 20- times, compared with the all virtualized simulation platform. Max. 24% performance improvement, 16% energy consumption reduction and 19% lifetime extension are achieved without requiring a cache layer for the SSD. Santa Clara, CA 12

13 Outline Introduction Three Techniques to Boost SSD Performance Storage Engine Assisted SSD (SEA-SSD) Logical Block Address (LBA) Scrambler Storage Class Memory (SCM)/ Flash Hybrid SSD Conclusion Santa Clara, CA 13

14 LBA scrambler-concept Actively write small data to fragmented pages (space utilization < 100%) so as to minimize the GC page-copy overhead. Free sector Sector with old data Sector with new data Blank page Fragmented page Invalid page Initial SSD state After writing data Conv. N valid : 3 Write new data... N valid : 3 Next erase block Currently N valid : 3 writing block N valid : 2... Prop.... Intentionally write new data to fragmented pages in the next erase block... Page Sector chip [1] A. Soga et al., IMW, May 2014, pp Santa Clara, CA [2] C. Sun et al., TVLSI, in-press. 14

15 (a) LBA scrambler in the host pros: smaller DRAM in SSD cons: interface modification required Conv. Prop. OS (File system) Host FTL Send LBA Send SLBA Translate LPA(SLPA) to PPA LBA Scrambler-System View OS (File system) LBA scrambler FTL Translate LBA into SLBAs of vacant sectors in the OP list* aggressively. Send SLPA in the next erase block to the OP list (Only write requests) (b) LBA scrambler in SSD pros: no need to modify the interface cons: Larger DRAM in SSD Conv. Prop. OS (File system) Host FTL Send LBA OS (File system) LBA scrambler Send SLBA Translate LPA(SLPA) to PPA FTL Translate LBA into SLBAs of vacant sectors in the OP list* aggressively. Send SLPA in the next erase block to the OP list (Only write requests) Send PA Send PA SSD SSD SSD SSD *Recommended writing pages are stored in the OP list *Recommended writing pages are stored in the OP list LBA: Logical block address SLBA: Scrambled LBA LPA: Logical page address SLPA: Scrambled LPA PA: Physical address PPA: Physical page address OP list : overwrite_preferred list FTL: Flash translation layer(ssd controller) [1] C. Sun et al., TVLSI, in-press. 15

16 LBA Scrambler-Algorithm Flow LBA Scrambler is based on the address-remapping with three rules. LBA scrambler Receive new write requests from host N int * (interval): LBA scrambler gets hint from FTL every N int writes Generate new write requests Y i = 0 i == N Int *? N i = i + 1 Enough SSD free space? N Y Rule (1) Write data to blank pages of the blocks Get recommended pages from FTL Generate new write requests Enough free space in next erase block? N Rule (3) Y Rule (2) Write to fragmented pages in the next erase block Send write requests to FTL Write to free space out of the next erase block Santa Clara, CA [1] C. Sun et al., TVLSI, in-press. 16

17 Short Summary LBA scrambler is able to improve the SSD write performance for all applications. With the LBA scrambler, small data are actively written to the remaining space of the fragmented page (the page utilization is less than 100%) in the next erase block of the SSD. Thus, the page-copy overhead is reduced when the GC is triggered. As a result, 35%-394% write speed boost, 27%-56% energy consumption reduction and 25%-55% endurance enhancement are achieved. Santa Clara, CA 17

18 Outline Introduction Three Techniques to Boost SSD Performance Storage Engine Assisted SSD (SEA-SSD) Logical Block Address (LBA) Scrambler Storage Class Memory (SCM)/ Flash Hybrid SSD Conclusion Santa Clara, CA 18

19 Memory overview Storage Class Memory (SCM) Mature Prototypical memory Emerging DRAM flash STT-MRAM PCM ReRAM Feature size (nm) <40 Cell area 6F 2 4F 2 20F 2 4F 2 4F 2 Read latency <10ns 0.1ms 35ns 12ns N/A W/E latency <10ns 0.1/1ms 35ns 100ns <10ns Retention 64 ms 10 year >10year >10year >=10year Endurance >1E16 1E5 >1E12 1E9 >1E6 W/E voltage (v) <3 Read voltage (v) <0.5 Single cell write energy (J/bit) 4E-15 4E E-12 6E-12 <=1E-12 Limited endurance [1] International Technology Roadmap for Semiconductors. Santa Clara, CA 19 SCM

20 SSD controller DRAM ReRAM flash memory SCM/ Flash Hybrid SSD A 3D TSV-integrated SCM/ flash hybrid SSD. DRAM is used to store management tables. TSV Host RAM Host interface Processor [1] H. Fujii et al., VLSI Circuits, 2012, pp Santa Clara, CA 20 [2] C. Sun et al., TCAS-I, vol. 62, no. 2, pp , FTL Addr. translation Wear leveling BBM&ECC Garbage collection SCM control unit SSD controller PCI Express, SATA, SAS MLC SLC SCM DRAM SSD

21 Data Storage in Hybrid SSD ReRAM supports in-place overwrite. Store fragmented/hot data to ReRAM. Store cold and seq. data to flash memory. Hybrid SSD (w/ CDE) Host MLC only Host SSD controller SSD controller ReRAM MLC Data Hot Cold Frag. Not frag. MLC All data store in MLC Cold data eviction frees ReRAM space for hot data Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 2, pp ,

22 SSD Data Management Algorithm A cold data eviction (CDE) algorithm is proposed as the replacement algorithm. Write new data N Evict cold and less fragmented data to Y Enough SCM free space? Page utilization small? Y Y Write new data to SCM N Page addr. is hot? N Cold & Un-fragmented Hot/Fragmented Data Data Write new data to Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 2, pp , END

23 SSD Write Performance Boost Over 10-times performance gain is achieved, compared with the flash only SSD. 128 SSD MLC total 64 capacity Hybrid SSD 32 (w/ CDE) 16 8 Hybrid SSD (w/o CDE) 4 only x12.6 X Write data size (GB) Write performance (MB/s) Santa Clara, CA [1] C. Sun et al., TCAS-I, vol. 62, no. 2, pp ,

24 Short Summary SCM is used as a storage device in SSD rather than a simple cache. Data placement strategy is to store fragmented data or hot data in SCM while cold and sequential data is stored in flash memory. It is achieved by the intelligent data management algorithms. For financial server application, over 10-times write performance boost can be achieved, compared with flash only SSD. TSV technology has been proved to be efficient of reducing the energy consumption of such hybrid SSD by over 90%. Santa Clara, CA 24

25 Outline Introduction Three Techniques to Boost SSD Performance Storage Engine Assisted SSD (SEA-SSD) Logical Block Address (LBA) Scrambler Storage Class Memory (SCM)/ Flash Hybrid SSD Conclusion Santa Clara, CA 25

26 Conclusion Three techniques have been proposed to improve the SSD write performance. SEA-SSD is proposed to improve the SSD write performance for the database application, by co-designing the database storage engine and SSD controller. LBA scrambler boosts the SSD write performance for general applications. It is based on the address-remapping technique. The middleware obtains the data storage information from the FTL. SCM/ flash hybrid SSD could own over 10-times faster write performance than the flash only SSD as a future hybrid solution. Santa Clara, CA 26

27 More Research SSD system: SSD intelligent data management (FTL) 3D-LSI circuit: boost converter-based adaptive voltage generator for 3D-SSD. Device: ReRAM, PCM, 0.5V low-power SRAM, Fe- flash etc. Santa Clara, CA 27

28 Thank you very much for your attention! This work is partially supported by NEDO. Santa Clara, CA 28