advances in flash memory technology & system … · ycsb benchmark set-up
TRANSCRIPT
1
Dr. John R. Busch
Advances in Flash Memory Technology & System Architecture to Achieve Savings in Data Center Power and TCO
Vice President and Senior Fellow October 18, 2013
2
Forward-Looking Statements
During our meeting today we may make forward-looking statements.
Any statement that refers to expectations, projections or other characterizations of future events or circumstances
is a forward-looking statement, including those relating to market position, market growth, product sales, industry
trends, supply chain, future memory technology, production capacity, production costs, technology transitions and
future products. This presentation contains information from third parties, which reflect their projections as of the
date of issuance.
Actual results may differ materially from those expressed in these forward-looking statements due to factors
detailed under the caption “Risk Factors” and elsewhere in the documents we file from time to time with the SEC,
including our annual and quarterly reports.
We undertake no obligation to update these forward-looking statements, which speak only as of the date hereof.
Cornell University – October 18, 2013
3
Overview
About SanDisk
Flash Trends
Flash Optimized Data Center Solutions
Conclusions
Cornell University – October 18, 2013
4
About SanDisk
5
A Global Leader in Flash Memory Storage Solutions
Financials as of Q2, ‘13. Net Cash = [Cash + cash equivalents + short-term & long-term marketable securities] less [debt at maturity value] as of the end of Q2, ‘13. Headcount & patents as of Aug., ‘13. NPD Estimate, Jan., ‘13. Estimates of the memory card & USB markets from NPD (Jan. ‘13) and GfK Retail and Technology, Oct., ‘12. Gartner: NAND Flash Supply & Demand, WW 1Q ‘12-4Q ‘14, 2Q ’13. Update Jun., ‘13.
The Leading Retail Brand in Key Markets
Close to Half of Industry Bit Output Together with manufacturing partner Toshiba
Technology Leadership
4,900+ Patents
1991 2013
Enterprise SSDs and Storage Software
Qualified at 6 of the Top 7
Server & Storage OEMs
All leading smartphone & tablet manufacturers use SanDisk
SanDisk Client SSD Design Wins at
11 Leading PC OEMs
The Leading Retail Brand in Key Markets
#1 Global Retail Revenue Share
Rankings Trailing 4 Qtr. Financials Global Operations Technology
5,000+ Employees
Fabs World Class NAND Capacity
19nm Leading Process Node
1Ynm Shipping
$5.6B Revenue
$4.4B Net Cash
$0.7B R&D Investment
Cornell University – October 18, 2013
6
Complete Top to Bottom Integration
Full Stack Enables Segment Optimized Solutions
Performance Scalability System
Utilization Endurance Cost Life Cycle
CONTROLLER NAND TECH NAND DIE WAFER SCALE MFG SSD SOFTWARE
Cornell University – October 18, 2013
7
Flash Trends
8
Bits per cell
Toshiba-SanDisk
~4F2 cell
Toshiba – SanDisk Flash Partnership
3D NAND & BiCS
Physical scaling: 210nm 160 130 90 70 56 43 32 24 19nm
Logical scaling: X1 (SLC) X2 (MLC) X3
Cornell University – October 18, 2013
BiCS = Bit Cost Scalable
9
2D-NAND Scaling Considerations
1Y technology node is 19/19.5 nm a substantial cell area reduction
1Y incorporates several new process modules to improve performance and reliability
1Z will leverage 1Y innovations to scale both X, Y dimensions substantial cell area reduction
19 nm 19/26nm
1Y 19/19.5nm
1Z Note: Diagram not to scale
Cornell University – October 18, 2013
10
3D NAND Technology BiCS – 3D charge trap structure Doesn’t require EUV Regular optical tools Bridge technology to 3D
ReRAM
Bit Lines
Source Lines
Select Gates
Word Lines
Back Gate
Y-cut: WL
X-cut: BL
Note: Diagram not to scale
3D NAND--Alternatives to Planar NAND
Cornell University – October 18, 2013
11
3D Resitor RAM : will follow 3D-NAND BiCS
3D ReRAM can scale to below 10nm node providing cost reduction beyond 2020
Cornell University – October 18, 2013
12
Spectrum of Memory Technologies
Cornell University – October 18, 2013
13
Positioning/Prospects of Memories
Write/Program Cycle Time (s)
Capa
city
(bits
)
1E-9 1E-8 1E-7 1E-6 1E-5 1E-4 1E-3 1E-2
1T
100G
10G
1G
100M
10M
1M
SRAM
DRAM
MRAM
PCM NOR
FG-NAND
BiCS
Code Storage
HDD
Working Memory
STT
ReRAM
Data Storage
ReRAM and BiCS are the two most promising post-2D NAND candidates
Progression of Memory Technologies
Cornell University – October 18, 2013
Source: SanDisk, presented at IMEC 2011
14
Flash Optimized Data Center Solutions
15
Flash is Enabling New Applications, Growing Fast
Source: Gartner
$24.4
$32.9
$13.2
$26.2
$38.7
$23.6
$28.5
$40.2 $38.3
$0
$10
$20
$30
$40
$50
DRAM HDD NAND
2008 2012 2016E
TAM ($B)
Cornell University – October 18, 2013
16
Flash Drives Savings in the Enterprise
Server/SAN Consolidation Hot, warm, and cold data Lower Cost SSD
Tota
l Co
st o
f O
wn
ersh
ip (
$)
High performance HDD
MLC Caching
In-Memory Computing Flash DIMM
3D Capacity HDD
SLC
Time
Cornell University – October 18, 2013
Cold Storage
17
Software Unlocks Flash Potential in the Enterprise Data Center
FLASH + SOFTWARE
New SAN Architecture
In-Memory Databases
Cold Storage
Big Data Analytics
Virtualization & Cloud Computing
Server-side Caching
Flash as replacement for 15k RPM HDD
18
Flash-Optimized Applications
Flash-optimized applications: – Exploit the high capacity, low latency, persistence and high throughput of flash memory
– Have extensive parallelism to enable many concurrent flash accesses for high throughput
– Use DRAM as a cache
– Get in-DRAM performance at in-flash capacity and cost, enabling server consolidation
Many applications realize limited benefits from flash without system level optimization
SanDisk ‘Flash Data Fabric (FDF)’ is a substrate for flash-optimized applications – Caching, key-value stores, databases, message queues, custom apps
– Leverages flash for high performance, high availability
– Enables low TCO through high server consolidation
– Executes on bare metal or virtualized
Cornell University – October 18, 2013
19
SanDisk ‘Flash Data Fabric (FDF)’ Enables Direct Flash Access for In-Memory performance
Hardware
Firmware
Driver
Operating System
Middleware
Applications
• Optimizes to fully exploit flash and multi-core
• Encapsulates optimizations for use by any application
• Open and standard initiative
Flash Data Fabric (FDF)
Cornell University – October 18, 2013
20
‘Flash Data Fabric’ features Provides an object API: create, replace, update, delete, indexes, range queries,
transactions, snapshot
Provides multiple namespaces via containers
Maps object keys to flash locations
Intelligent granular DRAM caching
Heavily optimized access paths for high performance
Optimized threading to maximize concurrency and minimize response time
Configurable flash management algorithms to optimize different workloads
Integrates with flash devices
– Minimizes write amplification, fast persistence, vectored operations,…
Executes in user space and is linked in as a dynamically loaded library
Cornell University – October 18, 2013
21
‘Flash Data Fabric’ Architecture
Container Mgmt Naming, create, open, delete
FDF Protocol Layer
Object Mgmt Naming, create, search,
update, delete,
Cluster Mgmt Naming, configure
Local DRAM
Caching
Flash Manager
Replication Elasticity Module
Messaging Subsystem Connect, send, receive
Transport Layer
Threading Module
Databases/Data Stores Data Grid and Object Stores
… Message Queue … … Session Store … Custom Apps
Cluster Services
Rep
licat
ion
Co
nfi
g
Failu
re H
and
ling
Fau
lt D
etec
tio
n
Application Layer
Cornell University – October 18, 2013
22
No SQL DataBase Example : Cassandra
Cassandra is an open source distributed key-value store
Key features:
– support for large scale synchronous and asynchronous replication, including across data centers
– automatic fault-tolerance and scaling
– tunable consistency (from “writes never fail” to “block for all replicas to be readable”)
– efficient support for large rows (1000’s of columns)
– CQL (SQL-like) query language
– supports multiple indices
FDF-Cassandra prototype based on Cassandra 2.1.4
Cornell University – October 18, 2013
23
Cassandra Performance 95/5 workload Stock Cassandra FDF Cassandra
Hard Drives 1.2k tps 100% HDD utilization 1 of 16 cores utilization
N/A
64GB Data (fits in memory)
40K tps 12 of 24 cores utilization
124K tps 18 of 24 cores utilization
256GB Data (data set in flash) 25K tps 90% flash utilization 18 of 24 cores utilization
95K tps 90% flash utilization 19 of 24 cores utilization
Intel Westmere server with 2 x 2.9GHz sockets, 24 cores, 96G DRAM
SSD: 8 x 200G SSD with software RAID 0
YCSB Benchmark set-up Remote client with 10G network connection
1K fixed object, uniform distribution with configurable read/write mix (eg: 95% read, 5 % update)
48GB FDF DRAM cache
Cornell University – October 18, 2013
24
TCO : Cassandra Requirement : 80k TPS and 1 TByte data set
$378,216
$55,620
$ 14,124
$10,000
$100,000
$1,000,000
Stock Cassandra on HDD stock Cassandra in DRAM FDF-Cassandra and Flash
TCO - Log Scale
3 Year OpEx
3 Year CapEx
Cornell University – October 18, 2013
Source: Based on internal testing
25
HDD DRAM SanDisk
No. of servers 34 6 1
Power (kW) 12.7 2.8 0.4
$ per transaction $8.44 $2.49 $0.51
2.4
80
95
0
10
20
30
40
50
60
70
80
90
100
HDD DRAM SanDisk
Tran
saconspersecond(Th
ousand)
Capacity
Performance
FDF + Flash Accelerates Database Performance at a Dramatically Lower TCO
Servers needed for 3.1TB dataset Yahoo! Cloud Serving Benchmark
95% Read 5% Write
Measurements on identical commodity x86 servers; scaled with modeling from 1.0 to 3.1TB Servers are running Apache Cassandra Database Management System
Cornell University – October 18, 2013
26
In Memory Data Grid Example: CouchBase vs FDF-Memcached
0
50000
100000
150000
200000
250000
300000
350000
400000
450000
500000
10
11
0
21
0
31
0
41
0
51
0
61
0
71
0
81
0
91
0
10
10
11
10
12
10
13
10
14
10
15
10
16
10
17
10
18
10
19
10
20
10
21
10
22
10
23
10
24
10
25
10
26
10
27
10
28
10
29
10
30
10
31
10
32
10
33
10
34
10
35
10
TPS
Couchbase
FDF-Memcached
Cornell University – October 18, 2013
27
In Memory Database Example: FDF-Redis Performance
116
84
93
70
93
132
101
114
99
89
0
20
40
60
80
100
120
140
String Hash List Set Sorted Set
KTP
S
Stock Redis (in memory)
FDF-Redis (out of memory)
FDF-Redis throughput with data set in Flash matches
Stock -Redis throughput with data set in DRAM
Cornell University – October 18, 2013
Bare Metal
Source: Based on internal testing
28
TCO : Stock Redis vs FDF-Redis Requirement : 80k TPS and 1 TByte data set
0
50,000
100,000
150,000
200,000
250,000
300,000
350,000
Stock-Redis AWS in DRAM FDF-Redis AWS with SSD
3 Year TCO
TCO
Cornell University – October 18, 2013
$-
$50,000
$100,000
$150,000
$200,000
$250,000
Stock Redis with DRAM96GB Servers
FDF-Redis with Flash
3 Year OpEx
3 Year CapEx
Bare Metal 3 Year TCO AWS
Source: Based on internal testing
29
Cloud Storage Example: FDF-Swift vs Stock-Swift Performance
Cornell University – October 18, 2013
5.8
93
149
222
2.628
40
80
134
2.635 17.261 19.672 19.493
2.632 9.852 10.431 9.242
0
50
100
150
200
250
1 8 16 32 64
TPS
in K
s
Clients
FDFSwift In-memory FDFSwift in-flash StockSwift in-memory StockSwift in-flash
Source: Based on internal testing, October 2013
30
Client Number
TPS in Ks FDF Object
server CPU
DRAM Miss rate
IO Utilization
1 2.6 95% 30% 5%
16 40 430% 58% 48%
32 80 1048% 58% 75%
64 134 2094% 58% 92%
Client # TPS in Ks Stock Object server CPU
IO Utilization
1 2 1000% 15%
8 9 2400% 45%
16 10 2400% 45%
32 9 2500% 45%
Stock-Swift in Flash
FDF-Swift in Flash
Cornell University – October 18, 2013
Source: Based on internal testing, October 2013
31
Conclusions
32
Conclusions
Flash technology trends will reduce TCO below all competing storage technologies
Many applications realize limited benefits from flash without optimization
Flash optimization of applications can yield near in-DRAM performance with balanced server with datasets residing in flash
Broad new set of flash use cases covering entire data center: hot to warm to cold data
,Cornell University – October 18, 2013
33
Thank you! © 2013 SanDisk Corporation. All rights reserved. SanDisk, SanDisk Ultra, SanDisk Extreme and SanDisk Extreme Pro are trademarks of SanDisk Corporation, registered in the United States and other countries. Lightning
is a U.S. registered trademark of SanDisk Enterprise IP LLC. iNAND Extreme is a trademark of SanDisk Corporation. ULLtraDIMM is a trademark of SanDisk Enterprise IP LLC. The SD and the SDHC mark and logo are
trademarks of SD-3C, LLC. Other brand names mentioned herein are for identification purposes only and may be the trademarks of their respective holder(s).
1GB=1,000,000,000 bytes. Actual user capacity less.