Addressing Scalability and Consistency Issues in Hybrid File System for BPRAM and NAND Flash

Transcription

1 7th IEEE International Workshop on Storage Network Architecture and Parallel I/O SNAPI 2011 Denver, Colorado May 25, 2011 Addressing Scalability and Consistency Issues in Hybrid File System for BPRAM and NAND Flash Embedded Software Systems Lab Dept of Electronics and Computer Engineering Hanyang University Quan Taizhong Jinsoo Yoo Jaemin Jung Youjip Won

2 Outline Background Structure of Compressed Metadata File System New Functionalities Experiment Conclusion 2

3 Advancement of Byte Addressable RAM(BPRAM) BPRAM Byte-addressable Non-volatile In-place update Faster than Flash memory ITEM DRAM FRAM PRAM MRAM NOR NAND Centering Byte Addressable YES YES YES YES Read only NO Non-volatile NO YES YES YES YES YES Read 10ns 70ns 68ns 35ns 85ns 15us Write 10ns 70ns 180ns 35ns 65us 200us Erase none none none none 700ms 2ms Power consumption High Low High Low High High Capacity High Low High Low High Very High Endurance 100K 100K Prototype Size 64Mbit 512Mbit 16Mbit 3

4 Issue for File System for NAND Flash NAND Flash DRAM In-memory Metadata Object Header Data Mount File Data SCAN Spare area SCAN Scan metadata from NAND Flash at mount time The mount time 4

5 Hierarchical File System for BPRAM and NAND Flash NEdel, DTuteja, Emiller, and SBrandt MRAMFS (2004, MASCOTS) SPark, TLee, and K, Chung RFFS (2006, ECS) JJung, YWon, EKim, HShin and BJeon FRASH (2010, TOS) IDoh, JChoi,DLee and SNoh MiNVFS (2007,ACM) YPark, SLim,CLee, and KPark PFFS (2008,ACM) AWang, PReiher, GPopek, and GKuenning The Conquest File System (2002, USENIX) MRAMFS RFFS FRASH MiNVFS PFFS The Conquest File System Hybrid system that stores small files and metadata in MRAM Stores the page data and page metadata separately Use BPRAM as storage of metadata in flash file system It uses Copy-On-Mount technique Maintains all the metadata in NVRAM Construct accessed small files in persistent RAMs double indirect index structure for data page management Storing metadata and frequently accessed small files in persistent RAMs double indirect index structure for data page management 5

6 Hierarchical File System for BPRAM and NAND Flash NAND Flash Object Header Data Data BPRAM BPRAM metadata BPRAM metadata BPRAM metadata BPRAM metadata Mount DRAM In-memory Metadata File COPY SCAN BPRAM area SCAN 6

7 Scalability Issue in Hierarchical File System Flash memory BPRAM CPM (Cost per MB) Low High Size ~ 64GB or more ~ 64MB Current Ratio 100 : 3 ( NAND : Data 512B, Metadata : 16B ) Needed Ratio 1000 : 1 7

8 Consistency Issue in Hierarchical File System 5sec cras h Unsynchronized FLASH Metadata W1 W2 W3 W4 W5 W6 mount procedure Time Periodic BPRAM update W : write request 8

9 Objective Hierarchical Filesystem for BPRAM and NAND Flash Scalability target : 1:1000 Instant Recovery for Unexpected Failure Compressed Metadata File System 9

10 Key Technical Ingredients Compressed Metadata File System BPRAM-aware Metadata Structure Hybrid Compression Algorithm DRAM Caching and periodic snapshotting Crash Recovery Scale down the BPRAM metadata Scale down the BPRAM metadata Reduce the Cost of BPRAM update Reduce Mount time After a crash Log Structured File System (YAFFS 20 Based) 10

11 System Architecture Processor cache Bridge/Memory controller memory BPRAM I/O bus NAND flash NAND flash 11

12 BPRAM-aware Meta Structure Object Header Data Data Block State Page Status ECC 64B ObjectID ChunkID ByteCount SerialNumber EC NAND Flash BPRAM CMFS_validitymarke r 8byte CMFS_device 40B + 8B * nblocks CMFS_object 24B + 36B * nobject SCM_vaild data_valid nerasedblock nfreechunck sequencenumber Block info structtype objectid parentid PAT 12

13 Hybrid Compression Algorithm in CMFS CMFS_validitymarker CMFS_device CMFS_object CMFS_object 24bytes + 36bytes * n field type field name size Not Compressed int structtype 4bytes unsigned int objectid 4bytes unsigned int parentid 4bytes LZO Compression int hdrchunk 4bytes enum varianttype:3 2bytes Huffman char Rice Shannonfano deleted:1 RLE LZO char softdeleted:1 char unlinked:1 char Hybrid Compression Algorithm char char fake:1 renameallowed:1 unlinkallowed:1 Huffman Rice Shannonfano 1byte RLE LZO char serial 4bytes int ndatachunks 4bytes array PAT (36bytes*n) 13

14 Target for Compression: PAT object PAT (Physical Address Translation) Tree Logical page number : level 1 file level 0 4 level 2 BPRAM level 3 Logical position 4byte Page numbers 32byte 14

15 Hybrid Compression high bit page number 16bit low bit Higher 8-bits are identical Lower 8-bits are unique 64 page pointers to a single block 16bit 0x00 0x61 0x00 0x62 0x00 0x63 0x00 0x64 0x00 0x65 Huffman compress No Compression 0x61 0x62 0x63 0x64 0x65 15

16 Coping with Size Variability in Compressed Metadata Link Based Object organization NAND flash In Memory ObjectID Invalid 262 Invalid 300 directory 262 file 300 Modify_objectID hd HASH TABLE Dir A s Object 262 File B s Object Sec BPRAM CMFS_device old metadata new metadata CMFS_object 16

17 Crash Recovery Minimizing Recovery Overhead: Using sequence number in the BPRAM NAND Flash block 100 BPRAM block 100 file file 1 Comparing Sequence number Object Object 3 Reconstruct BPRAM metadata block 102 block 101 block 102 file file 2 Scanning NAND Metadata Object Object Object Object 17

18 Detecting Inconsistency Comparing sequence number Block information in CMFS_device : Empty Block information in NAND flash : Being allocated 100 Scan & Rebuild Full Compar sequence number 18

19 Experiment Experimental Environment Marvell PXA320 Board 806 MHz 128 MB SDRAM 128 MB large block NAND flash memory Linux kernel 2624 A portion of DRAM is used to emulate BPRAM 19

20 Experiment: Hybrid Compression When the file size > 16KB, Hybrid compression 31% better than LZO 20

21 Experiment: Metadata Update(LMBEMCH) - 202% - 233% File Creation File Deletion 21

22 Experiment: Random/Sequential Write(IOZONE) No performance hit even with compression and metadata synch operation % % 22

23 Experiment: Mount Overhead YAFFS : the mount latency linearly increases CMFS : the mount latency does not vary subject to the change in file system partition size 23

24 Conclusion CMFS is a novel hierarchical file system which provides efficient management method of BPRAM BPRAM-aware metadata design for scalability Hybrid compression for scalability BPRAM update method to reduce the Cost of BPRAM update Efficient Detection of Inconsistent Blocks and Crash recovery 24

25 The end Thank You 25