Replacing the FTL with Cooperative Flash Management

Transcription

1 Replacing the FTL with Cooperative Flash Management Mike Jadon Radian Memory Systems Flash Memory Summit 2015 Santa Clara, CA 1

2 Data Center Primary Storage WORM General Purpose RDBMS HDD Flash DRAM Flash Memory Summit 2015 Santa Clara, CA 2

3 Data Center Primary Storage WORM General Purpose RDBMS HDD Flash DRAM Flash Memory Summit 2015 Santa Clara, CA 3

4 Unpredictable Latency Spikes (QoS) Cost Endurance (Wear Out/TCO) Flash SSD

5 Advanced Systems Purpose-Built storage systems All-Flash Arrays Hybrid SSD/HDD Advanced local file systems Software-Defined Storage architectures Hyperconverged systems Object/Key Value Stores Certain In-Memory Data Base applications

6

7 Backwards Compatible Simple Contract

8

9 IOPS

10

11

12 Standardizing Storage Intelligence for SSDs TRIM Hinting and Optimal Stripe Size FTL SSD Intelligence Multi-Stream Writes Intelligent placement of data on the storage device Intelligent management of garbage collection and overprovisioning

13 Challenges with FTL Storage Intelligence

14 Challenges with FTL Storage Intelligence Universal API Huge number of potential parameters Information on caching, prioritization, garbage collection, scrubbing, hot/cold Information has to be continuously prepared and communicated on every op Set up overhead on system software and through communication chain Complex Host/Device Contract

15 When and what function will be offered by each vendor? Will need to continuously add options Each vendor s implementation will produce different results, interoperability, and dependencies

16 Data Lifecycle in Storage Software Management is required across the life cycle of data Process alignment

17 Data Lifecycle in Storage Software FTLs duplicate these processes High probability for disconnects Performance and latency impacted with every disconnect

18

19

20 SSD Storage Intelligence and Latency Spikes Read collisions Your Read latency is your Write latency!

21 SSD Storage Intelligence and Latency Spikes

22 Lifecycle Scheduling and Execution Proprietary I/O pipeline Internal system with async processes: cache flush, garbage collection, data refresh, etc. Complex controls to mitigate

23 Lifecycle Scheduling and Execution

24 Software-Defined Flash Directly Export Physical Geometry of NAND array Directly Export BBM/ECC handled on device BBM/ECC Software-Defined Flash: No FTL Direct Geometry Export Host stack performs all Flash Management processes Results: Enables maximum parallelization Dramatic performance improvements

25 Software-Defined Flash Directly Export Physical Geometry of NAND array Directly Export BBM/ECC handled on device BBM/ECC Challenges: Integration Burdens system with NAND attributes and constraints Does not provide Forward Compatibility Poor RAS capabilities Poor Scalability

26 Flash is now the most important Storage System media Flash SSD But every Flash SSD s software is designed to emulate a Hard Disk Drive Hard Disk Drive

27 Flash is now the most important Storage System media Flash SSD NVMe or SAS Cooperative Flash Management Conventional SSD Hardware

28 System-Driven Architecture

29 Offload Engine

30 Backwards Compatible Simple Contract

31 Offload Engine

32 Offload Engine 75% improvement in System-Level Write Amplification

33 Scalability

34 Abstraction Geometry Emulation Address Configurator Enables Forward Compatibility Reliability and FRU capabilities Vendor Supported Warranties

35 Solving Software-Defined Flash for the Data Center Directly Export ECC SDF Challenges: Integration Burdens system Forward Compatibility RAS Scalability CFM: Geometry Emulation & Address Configurator Offload Engine Wear Leveling

36 Extensibility & Parallelization

37 Data Lifecycle in Storage Software FTLs duplicate these processes High probability for disconnects Performance and latency impacted with every disconnect

38 Data Lifecycle in Storage Software CFM enables host ownership of the complete lifecycle in a combined system Implicitly supports different storage stacks with unique implementations and requirements

39 Lifecycle: Scheduling and Execution CFM enables host optimization of flash resources High Visibility into I/O queuing High Parallelization Predictable operation Low-Latency & Low-Jitter I/O pipeline proprietary Internal storage system Internal Async processes (cache flush, gc, data refresh, etc) Complex controls to mitigate

40 Lifecycle: Scheduling and Execution Is R2 data on D0 or D1? When does the GC submit access the flash dies? When does the cache flush?

41 Lifecycle: Scheduling and Execution System leverages workload knowledge and CFM visibility With CFM, the host can see W1 will access same die as R2 and take action: e.g. prioritize R2

42 CFM Host Control System intelligence is left in the host and preserved to the media, so it s independent of the SSD Simple host contract from the outset not a complex evolving interface

43 Extensibility & Parallelization

44 Workload: 66% random reads/33% sequential writes on random 8MB segments Device Queue Depth: 128; Assumes 1.5 Host WA factor >80% improvement in IOPS and Bandwidth Symphonic firmware turns the SSD into an offload engine, eliminating the copying that would normally occur between the host and SSD (copy overhead that would even occur with a HDD).

45 Cooperative Flash Management

46 Extremes FTL FTL SSD True SDF Flash Memory Summit 2015 Santa Clara, CA 46

47 Goldilocks FTL FTL SSD CFM SSD True SDF Flash Memory Summit 2015 Santa Clara, CA 47

48 Flash is now the most important Storage System media 2015 Where does Flash go from here?

49 Mean Avg. Mean Avg. Mean Avg.

50

51

52

53

54 Data Center Primary Storage HDD Flash DRAM Flash Memory Summit 2015 Santa Clara, CA 54

55 Data Center Primary Storage HDD Flash DRAM Flash Memory Summit 2015 Santa Clara, CA 55

56 WORM Data Center Primary Storage General Purpose RDBMS HDD FTL Flash DRAM Flash Memory Summit 2015 Santa Clara, CA 56

57 Data Center Primary Storage WORM General Purpose RDBMS HDD FTL Flash CFM Flash DRAM Flash Memory Summit 2015 Santa Clara, CA 57