Professional Documents
Culture Documents
Solid-State Storage:
Technology, Design and
Applications
Abstract
Author Biographies
Rich Freitas is a Research Staff Member at the IBM Almaden Research
Center. Dr. Freitas received his PhD in EECS from the University of
California at Berkeley in 1976. He then joined IBM at the IBM T.J.
Watson Research Lab. He has held various management and research
positions in architecture and design for storage systems, servers,
workstations, and speech recognition hardware at the IBM Almaden
Research Center and the IBM T.J. Watson Research Center. His current
interest lies in exploring the use of emerging nonvolatile solid state
memory technology in storage systems for commercial and scientific
computing.
Acknowledgements
Winfried Wilcke
Geoff Burr
Bulent Kurdi
Clem Dickey
Paul Muench
C. Mohan
KK Rao
Agenda
Introduction 10 min
Technology 40 min
System 30 min
Questions 10 min
Break 30 min
Applications 40 min
Performance 40 min
Questions 10 min
Introduction
Speed/Volatility/Persistence Matrix
HDDs Magnetic
head
(slider)
$ High growth in disk areal density has driven the HDD success
$ Magnetic thin-film head wafers have very few critical elements per
chip (vs. billions of transistors per semiconductor chip)
$ Thin-film head (GMR-head) has only one critical feature size
controlled by optical lithography (determining track width)
$ Areal density is control by track width times (X) linear density
Solid State Storage: Technology, Design and Applications 2010
9 FAST February 2010 IBM Corporation
Future of HDD
Higher densities through
perpendicular recording
Jul 2008
610 Gb/in2 ~4 TB
patterned media
www.hitachigst.com/hdd/research/images/
pm images/conventional_pattern_media.pdf
Bandwidth
[MB/s]
100
100
MB/s
10
10
1
1985
1985 1990
1990 1995
1995 2000
2000 2005
2005 2010
2010
Date Available
Date
Latency 6
[ms]
5
0
1985 1990 1995 2000 2005 2010
Date
Bandwidth Problem is getting much harder to hide with parallelism
Access Time Problem is also not improving with caching tricks
Power/Space/Performance Cost
Solid State Storage: Technology, Design and Applications 2010
13 FAST February 2010 IBM Corporation
2
similar understanding for Flash & other
0
SCM technologies not yet available 1 2 3 4 5
age of drive
Consider: drive failures during 12
AFR
recovery from a drive failure? 10
[%]
8 by usage
6
4
potential for improvement given
2
switch to solid-state (no moving parts)
faster time-to-fill (during recovery) 0
1 2 3 4 5
[Pinheiro:2007] age of drive
Solid State Storage: Technology, Design and Applications 2010
14 FAST February 2010 IBM Corporation
IBM Almaden Research center
Extrapolation to 2020
(at 70% CGR need
2 GIOP/sec)
5 million HDD
16,500 sq. ft. !!
22 Megawatts
R. Freitas and W. Wilcke, Storage Class Memory: the next storage
system technology to appear in "Storage Technologies & Systems"
special issue of the IBM Journal of R&D.
Solid State Storage: Technology, Design and Applications 2010
15 FAST February 2010 IBM Corporation
Billions!
Memory-type
Uses
Storage-type
Uses
Power!
Solid State Storage: Technology, Design and Applications 2010
17 FAST February 2010 IBM Corporation
Technology
64Mb FRAM (Prototype) 4Mb MRAM (Product) 512Mb PRAM (Prototype) 4Mb C-RAM (Product)
0.13um 3.3V 0.18um 3.3V 0.1um 1.8V 0.25um 3.3V
organic
50
solid electrolyte
RRAM
40 PCram
MRAM
30 FeRAM
SONOS Flash
nanocrystal Flash
20
Flash
10
0
2001 2002 2003 2004 2005 2006 2007 2008
Industry interest
in non-volatile memory 2001
2006
www.itrs.net
Density is key
Cost competition between
IC, magnetic and optical
2F 2F
devices comes down to
effective areal density.
What is Flash?
oxide
gate
source drain Based on MOS transistor
NOR NAND
Cell Size 9-11 F2 2 F2
(4 F2 physical x 2-bit MLC)
Published
ITRS 2003
CMOS Logic CMOS Logic
TaN
oxide/nitride/oxide
oxide
Evolution??
Migrating to Semi-spherical
TANOS memory cell 2009
Migrating to 3-bit cell in 2010
Migrating to 4-bit cell in 2013
Migrating to 450mm wafer size in
Chip Density, GBytes
2015
Migrating to 3D Surround-Gate Cell
in 2017
Source: Chung Lam, IBM
Solid State Storage: Technology, Design and Applications 2010
32 FAST February 2010 IBM Corporation
IBM Almaden Research center
Saturation charge
What is FeRAM? Remanent
charge
ferroelectric material
Qs
such as
lead zirconate titanate +
+
+
+
+
+ Qr
- - -
(Pb(ZrxTi1-x)O) or PZT - - -
metallic electrodes VC
-Qr Coercive
-Qs voltage
need select transistor
half-select perturbs
perovskites (ABO3) = 1 family of FE materials
destructive read forces need
for high write endurance
inherently fast, low-power, low-voltage
[Sheikholeslami:2000]
Solid State Storage: Technology, Design and Applications 2010
33 FAST February 2010 IBM Corporation
[Gallagher:2006]
Solid State Storage: Technology, Design and Applications 2010
34 FAST February 2010 IBM Corporation
IBM Almaden Research center
Memristor
Memristor what is
Note time-range chosen for simulations,
and the required switching current (power)
scalebar?
Solid Electrolyte ON
Resistance contrast by forming a
state
metallic filament through insulator
sandwiched between an inert
cathode & an oxidizable anode. OFF
state
Ag and/or Cu-doped
GexSe1-x, GexS1-x or GexTe1-x
Cu-doped MoOx [Kozicki:2005]
Cu-doped WOx
RbAg4I5 system
Advantages Issues
Program and erase at very low Retention
voltages & currents
Over-writing of the filament
High speed
Sensitivity to processing temperatures
Large ON/OFF contrast (for GeSe, < 200oC)
Good endurance demonstrated Fab-unfriendly materials (Ag)
Integrated cells demonstrated
Solid State Storage: Technology, Design and Applications 2010
40 FAST February 2010 IBM Corporation
IBM Almaden Research center
Amorphous phase
High resistance
Low reflectivity
[Neale:2001]
[Wuttig:2007]
Solid State Storage: Technology, Design and Applications 2010
41 FAST February 2010 IBM Corporation
Phase-Change Bit-line
RAM PCRAM
programmable
resistor
Potential headache:
Voltage RESET pulse High power/current
temperature Tmelt affects scaling!
Tcryst
SET Potential headache:
pulse If crystallization is slow
time affects performance!
temperature
Tmelt
How a phase-change cell works Tcryst
crystalline amorphous SET
state state
pulse time
S. Lai, Intel,
IEDM 2003.
heater
wire
access device
Mushroom cell
SET state
LOW resistance
RESET state
HIGH resistance
Tmelt
How a phase-change cell works Tcryst
crystalline SET
state amorphous
state
pulse time
S. Lai, Intel,
IEDM 2003.
heater
wire
access device
Vth
SET state
LOW resistance
RESET state
HIGH resistance
Hold at
slightly under Filament
melting during Field-induced
broadens, electrical
recrystallization then breakdown
heats up starts at Vth
Solid State Storage: Technology, Design and Applications 2010
44 FAST February 2010 IBM Corporation
IBM Almaden Research center RESET pulse
temperature
Tmelt
How a phase-change cell works Tcryst
crystalline amorphous SET
state state pulse time
S. Lai, Intel,
IEDM 2003.
heater
wire
access device
SET state
LOW resistance
RESET state
HIGH resistance
Issues for phase-change memory
Keeping the RESET current low
Multi-level cells (for >1bit / cell)
Is the technology scalable?
Solid State Storage: Technology, Design and Applications 2010
45 FAST February 2010 IBM Corporation
Scalability of PCM
Basic requirements
widely separated SET and RESET resistance distributions
switching with accessible electrical pulses
the ability to read/sense the resistance states without perturbing them
high write endurance (many switching cycles between SET and RESET)
long data retention (10-year data lifetime at some elevated temperature)
avoid unintended re-crystallization
fast SET speed
MLC capability more than one bit per cell
Phase-Change Nano-Bridge
Prototype memory device with ultra-thin (3nm) films Dec 2006
2F 2F add N 1-D
starting sub-
demonstrated
from lithographic (at IEDM 2005)
standard fins (N2 with
4F2 store M bits/cell 2-D)
with 2M multiple
levels go to 3-D with
L layers
demonstrated
(at IEDM 2007)
Sub-lithographic addressing
Push beyond the lithography
roadmap to pattern
a dense memory
[Gopalakrishnan:2005 IEDM]
Solid State Storage: Technology, Design and Applications 2010
49 FAST February 2010 IBM Corporation
16k-cell
array
[Nirschl:2007]
Solid State Storage: Technology, Design and Applications 2010
51 FAST February 2010 IBM Corporation
3-D stacking
Stack multiple layers of memory
above the silicon in the
CMOS back-end
SE M. N. Kozicki, M. Park, and M. Mitkova, IEEE Trans. Nanotech., 4(3), 331-338 (2005).
M.N. Kozicki, M. Balakrishnan, et. al., Proc. IEEE NVSM Workshop, 83-89 (2005).
M. Kund, G. Beitel, et. al., IEDM Tech. Dig., 754-757 (2005).
P. Schrgmeier, M. Angerbauer, et. al., Symp. VLSI Circ., 186-187 (2007).
FeRAM/MRAM/RRAM/SE References
In comparison
SONOS Nanocrystal
Flash Flash Flash FeRAM FeFET
advanced
Knowledge level product development product basic research
development
low switching
yes yes yes yes yes
Power
high poor
endurance
NO NO yes yes
(1e7 cycles)
Comparison continued
solid organic
MRAM Racetrack PCRAM RRAM electrolyte memory
basic advanced Early
Knowledge level product development basic research
research development development
high
endurance
yes should yes poor unknown poor
Systems
109
Storage
108
Year
107 Access DISK (5ms)
Human Scale
Month 106
105
Time in ns
FLASH
2008 CPU RAM DISK TAPE
SSD
RAM
Architecture
Synchronous
Internal Hardware managed
CPU Low overhead
DRAM
Processor waits
Memory Fast SCM, Not Flash
Controller Cached or pooled memory
I/O SCM
Controller
SCM Asynchronous
Software managed
High overhead
Processor doesnt wait
Switch processes
SCM
Storage Flash and slow SCM
Controller Paging or storage
Disk
External
Power 100mW
Page
Block
B
U Page
Data erased in blocks UCH
F Page
Block = 64 - 128 Pages FE Page
Page
Block
Page
Page
Page
Block
20 MB/s transfer rate sustained
ONFI will take it to 200 MB/s
Page
Write streams data from host into
Page
BUF Page
6 MB/s transfer rate sustained Page
BCA
Block
20 MB/s on standard bus B
U Page
ONFI increases this to F
UCH Page
Program copies BUF into an erased FE Page
Page
Program 2 KB / 4 KB page: 0.2 ms Page
Block
Erase clears all Pages in a Block to Page
1s Page
Erase 128 KB block: 1.5 ms Page
A block must be erased before any of
its pages may be programmed
CPU MEMORY
This can be the basis for a card design, an SSD design or a subsystem design.
$10,000.00
$1,000.00
$100.00
$10.00
DRAM
$1.00 Flash
SRAM
MRAM
$0.10 DRAM-pr
FLASH-pr
SCM
$0.01
1990 1995 2000 2005 2010 2015 2020
Input
Priceform the
Trends: Subsystem
Magnetic Price
disks and Crystal
Solid State Ball
Disks
$100.00
Ent Flash SSD
PCM SSD
Ent. Disk
$10.00
SATA Disk
Price: $/GB
$1.00
$0.10
SSD Price = multiple of device cost
2
10x SLC @ -40% CAGR (4.5F )
2
3x MLC @ -40% CAGR (2.3 --> 1.1F )
2
3x PCM @ -40% CAGR (1.0 F )
$0.01
2004 2006 2008 2010 2012 2014 2016 2018 2020
40
MB/s
1000
17
10 7 100 49
10
1 USB disk LapTop Enterprise
USB disk LapTop Enterprise
Sustained Read Bandwidth Sustained Write Bandwidth Maximum Random Read IOPs Maximum Random Write IOPs
Logical Address > Translation > Wear Level > SCM Physical Add
DRAM
X86, PowerPC, DIMM
SPARC
CPU VM PMC
DRAM
X86, PowerPC, DIMM
SPARC
PMC
CPU VM
SCMC
SCM
Card
SCM SCM
Controller
Solid State Storage: Technology, Design and Applications 2010
77 FAST February 2010 IBM Corporation
X86, PowerPC,
SPARC
CPU VM PMC
SCM
Card
SCM Controller
Wear Leveler SCM
Summary
There are a number of solid state memory
technologies competing with DRAM and Disk
Flash and PCM are the current leaders
An inexpensive nonvolatile memory with medium
speed (1 50 us) will change the storage hierarchy
An inexpensive memory with speed near DRAM will
change the memory hierarchy
Such a memory that is also nonvolatile will enable new
areas
Write endurance is an issue for many of these
technologies, but there are techniques to cope with it
Questions?
Break Time
Performance
IO Read -> Flash Read, IO Write -> Flash Erase and Flash Program
Minimal Latency
R/W IOPS Latency Queue
- usecs Depth
100/0 17221 56.9 1
0/100 11316 85.8 1
50/50 11776 83 1
Minimum Latency
R/W IOPS Latency Queue
- usecs Depth
Pre-conditioning impacts performance Lose 5-10K IOPS per 10% of write ratio
Developed IO stream
shaping algorithm to
improve SSD R/W Bathtub
Performance behavior.
4K 16K 64K
Random IOPS
100R 1526330 758593 190898
50/50 RW 1087675 345763 116967
100W 500636 177767 58606
IOPS
IO Block Size
Application
Smart
Virtual volumes
Physical Devices
Workload Learning
Each workload has its unique IO
Time (Minutes) access patterns and characteristics
over time.
Latency
IO Density
Each circle represents an IO (density,latency) point. The size of the
circle is proportional to the number of IOs at that point. Color simply
corresponds to location in the coordinate space; it has no additional
meaning. Each circle has the same opacity.
Solid State Storage: Technology, Design and Applications 2010
107 FAST February 2010 IBM Corporation
Data Placement
Setup:
60-70%+
60-70%+Reduction in "SPC-1inLike"
Reduction Average
SPC-1 Response
Like Time
Average
Single Enterprise Storage System with
Response with
Time withSmart
DS8000 DataTiering
Placement Technology
Technology both HDD and SSD ranks. About 5-6%
capacity is in SSD ranks.
10 100.00%
HDD only
R e s p o n s e T im e (m s )
0 0.00% Result:
0 5000 10000 15000 Response time reduction of 60-70+% at
peak load
IOPS Sustainability test, 76%
Ramp test, 77%
Migration Begins
after 1 hour
HDD+SSD
with Data
Placement
Double IOPS with
Constant latency
Scalable Throughput
300% throughput
Transaction Per Second
improvement.
300%
Throughput
Improvement
Substantial IO Bound
Transaction Response
time Improvement
45%-75% response HDD Only
time improvement.
Architecture
Synchronous
Internal Hardware managed
CPU Low overhead
DRAM
Processor waits
Memory Fast SCM, Not Flash
Controller Cached or pooled memory
I/O SCM
Controller
SCM Asynchronous
Software managed
High overhead
Processor doesnt wait
Switch processes
SCM
Storage Flash and slow SCM
Controller Paging or storage
Disk
External
Memory
Possible positioning in the memory stack
Paging
Summary
SCM in the form of Flash and PCM are here today
and real. Others will follow.
Q&A