Slides(PPT) - Storage Research Group

Report
Extending the Lifetime of Flash-based Storage
through Reducing
Write Amplification from File Systems
Youyou Lu, Jiwu Shu, Weimin Zheng
Tsinghua University
Outline
•
•
•
•
•
Background and Motivation
Object-based Flash Translation Layer
System Co-design with Flash Memory
Evaluation
Conclusion
Flash Memory
• Gained Popularity
– High performance, low energy, reduced cost
– Wide deployment: embedded devices ->
enterprise servers
• Endurance
– SLC (100,000)
– MLC (10,000)
– TLC (1,000)
Figure from “The Bleak Future of NAND Flash
Memory” in FAST’12
Existing Approach to Flash Endurance
• Wear Leveling
– To make all the blocks evenly worn out
– Fundamental part of the FTL
• Data Reduction
– To reduce the amount of data to be written
– Data De-duplication and Compression
– Used either in FTL or in FS
Write Amplification from File Systems
• Write Amplification from File Systems
– Pre-FS writes vs. Post-FS writes
• Journaling
– Keep the journals in the logs first,
– And then, checkpoint them in-place
• Metadata synchronization
– Frequent persistence in case of data lost or
inconsistency
• Page-aligned update
– Wasted space within one page
A simple example in ext3
• Echo “title” > foo.txt
– Effective Data: 6 bytes
– Flash Writes: 11 pages * 4KB/page = 44KB
• Echo “texttexttext…”(4KB) >> foo.txt
– Effective Data: 4KB
– Flash Writes: 9 pages * 4KB/page = 36KB
bmp
data
bmp
inode
Data Area
data
dirent
bmp data
dirent
C
data inode
bmp
bmp
C
Journal Area
inode
data
Flash Opportunities
• No-overwrite
– Can the journaling use it without writing twice in
the file system?
• Page metadata
– Can we store the backpointer to lazily write back
the index while keeping consistency?
• Erase-before-Update
– Can we track the free space in a coarse-grained
way?
Outline
•
•
•
•
•
Background and Motivation
Object-based Flash Translation Layer
System Co-design with Flash Memory
Evaluation
Conclusion
Architecture & Data Organization
• Intelligent Storage Mgmt.
• Optimized FS Mechanism
Application
VFS Syscalls
File System
File System
(Namespace Mgmt.
+
Storage Mgmt.)
VFS Syscalls
Application
Namespace Mgmt.
Object File System
OBJ_READ, OBJ_WRITE, OBJ_FLUSH, OBJ_CREATE,
OBJ_DELETE, OBJ_SETATTR, OBJ_GETATTR
Storage Mgmt.
Object-based Flash Translation Layer (OFTL)
READ, WRITE
FLASH_READ, FLASH_WRITE,
FLASH_ERASE
Flash Translation Layer (FTL)
Solid State Drive
Hardware
Raw Flash Device
Hardware
OFTL Data Layout
Root Page
Object Index
MAGIC NUMBER
ADDR_BLK_INFO
ADDR_ATIME_LOGS
ADDR_OBJ_INDEX
ADDR_UW_META
Block Info Meta
Block Info Segment
Object Metadata
File Metadata
Layout
Diff-Layout
Object Metadata
File Metadata
Layout
Diff-Layout
Diff-Data
Diff-Data
Object
Metadata
Pages
Block Info Segment
Block Info Segment
Coarse-grained Block
Lazy Indexing
State Maintenance
Object
Data
Pages
Compacted Update
T1. Lazy Indexing
• Index Metadata
– The metadata that stores the pointers (the physical
addresses of other pages).
• Object Index -> Object Metadata Page -> Object
Data Page
– Type-specific backpointers
Page Metadata
(OOB)
PAGE DATA
TYPE
BACK-POINTER
(oid, offset, len)
Ver
Tid
Tcnt
...
ECC
Updating Window & Checkpointing
Updating Window
Updating Window
1.Make sure the mappings are persistent
2.Write back the updating window metadata
• Transactional write
– <tid, tcnt>
– Count the total number of the pages with the
same tid, and compare with the stored tcnt
Page Metadata
(OOB)
PAGE DATA
TYPE
BACK-POINTER
(oid, offset, len)
Ver
Tid
Tcnt
...
ECC
T2. Coarse-grained Block State
Maintenance
• Free space in flash block units
– Page states can be identified using the block state
• Pages in FREE blocks are all free
• Pages in USED blocks are all used
• Pages in UPDATING blocks need to be further identified
• Relaxed Metadata Persistence
UPDATING
w )
do n
in atio
c
g W lo
in (Al
t
da ing
Up nd
te
Ex
FREE
Erase
U
Ex p d
te at
nd in
in g W
g
(E ind
vic ow
tio
n)
USED
T3. Compacted Update
• Compact multiple partial page updates into
one flash page
• Co-locate the diff-page with the metadata
page
Diff-Layout
Diff-Extent
(3712, 384, 0)
o_off
Diff-Extent
(4096, 768, 384)
len
addr
Diff-Extent
Diff-Extent
(8192, 896, 1152) (16384, 196, 2048)
Diff-Page
0
511
Diff-Data (0)
Diff-Data (1)
Diff-Data (2)
Diff-Data (2)
Diff-Data (3)
An Example in OFSS
[Inode], [diff-data: “title”]
[data: “title texttexttext…”]
[dirent: “foo.txt”]
[Inode], [diff-data: “xttext”]
OMP, <oid2, 0, 0>, ver1,
Tx1, 0, [ECC]
ODP, <oid2, 0, 4096>,
ver3,
Tx2,
0, [ECC]
ODP,
<oid1,
0, 4096>,
ver2, Tx1, 2, [ECC]
OMP, <oid2, 0, 0>, ver4,
Tx2, 2, [ECC]
Updating Window
• Echo “title” > foo.txt
• Echo “texttexttext…”(4KB) >> foo.txt
Comparison of Ext3 and OFSS
•
•
•
•
80 KB (ext3) -> 16 KB (OFSS)
Journals => Transactional Metadata in Page
Metadata
Inode => Reverse Index in Page Metadata
Block/Inode Bitmap => Free Space Mgmt. in
Flash Block Units
Page Un-aligned Update => Compaction and
Co-location
Outline
•
•
•
•
•
Background and Motivation
Object-based Flash Translation Layer
System Co-design with Flash Memory
Evaluation
Conclusion
Evaluation Method
• Evaluation Metric
– write amplification = flash_writes / app_writes
• System Evaluation Framework
– OFSS
– File system
RWC Trace
Object File
System
OFTL
RWC Trace
File System
Block Trace
Utility
Block Trace
PFTL
Flash R/W/E
Flash R/W/E
Overall Comparison (1)
Write Amplification:
OFSS = 15.1% * ext3 = 52.6% * ext2 = 10.6% * btrfs
Overall Comparison (2)
Write Amplification:
OFSS = 36.0% * ext3 = 80.2% * ext2 = 51.0% * btrfs
Metadata Amplification
In OFSS, meta
Ext3
ampl. is
Ext2
dramatically
OFSS
reduced
Ext3:
Journaling
Ext2: Bitmap
and Inode
Table
Refer to the paper
for more details of
the async mode
iPhoto
iPages
LASR1
LASR2
LASR3
TPCC
2.57
1.06
0.05
8.59
2.68
0.30
17.91 11.84 51.04 3.73
2.00 0.91 4.11 1.04
1.13 0.45 1.05 0.03
Flash Page Size Impact
• Write amplification gets worse and worse as the flash
page size increases
• The sync mode is much more worse than the async
Refer to the paper for more details of the async mode
Outline
•
•
•
•
•
Background and Motivation
Object-based Flash Translation Layer
System Co-design with Flash Memory
Evaluation
Conclusion
Conclusion
• Metadata in file systems are frequently
written back for consistency and durability,
amplifying the writes to the flash memory
• Flash memory offers opportunities for
endurance-aware file system mechanisms
– Journaling: transactional write
– Metadata Synchronization: lazy indexing, coarsegrained block state maintenance
– Page-aligned Update: compacted update
Thanks!
Questions?
Youyou Lu, Jiwu Shu, Weimin Zheng
(luyy09@mails.tsinghua.edu.cn)
http://storage.cs.tsinghua.edu.cn/~lu
Backup – ext3 layout
Block Group
Super
Block
Group
Desp.
Block Group
Inode
Bitmap
Block
Bitmap
Block Group
Inode
Table
Data Blocks
Backup – Metadata Amplification (async)
Backup – Impact of Flash Page Size (async)
Backup – Overhead of Window Extending

similar documents