SSD

Performance and Reliability
Improvement of SSDs
Beom Ju Shin
SK hynix
October 7, 2015
Presenter
• 1999 ~ 2007: DRAM
− DDR SDRAM design
− DDR3 SDRAM design
− GDDR5 SDRAM design
• 2008 ~ : NAND flash memory
Beom Ju Shin
Senior Engineer / PL
SK hynix
− Triple-level cell (TLC) design
− Perfect page NAND (PPN) design
− SSD performance analysis
2
NVM on Memory Hierarchy
• Non-volatile memory (NVM), such as SSD, is
filling the price/performance gap between DRAM
and HDD.
SSD
Source: Amber Huffman, Flash Memory Summit 2010
3
Linux I/O Stack
• Different layers provide a single, virtual interface
to the respective layers above.
• SSD is located below the I/O stack.
Applications
Virtual file system (VFS)
Block based
FS
Network FS
Pseudo FS
Special
purpose FS
Direct I/O
Page cache
Block I/O layer
I/O scheduler
SCSI layer
HDD / SSD
4
Data Storage Market Trend
• For PCs, SSD shipments will continue to increase
while HDD shipments will decline in the future.
Worldwide shipments for HDDs and SSDs in PCs
500M
400M
300M
200M
SSD
100M
HDD
0M
2012
2013
2014
2015
2016
2017
Year
Source: IHS Inc., 2013
5
HDD vs. SSD
• SSD contains no mechanical parts and consists
of a few major components.
HDD
SSD
Source: Inside NAND Flash Memories
6
HDD vs. SSD (Cont.)
• One SSD can replace several HDDs as it contains
multiple NAND flash memory chips and can
operate them in parallel.
Source: Inside NAND Flash Memories
7
HDD vs. SSD (Cont.)
• SSDs are superior to HDDs in all aspects except
for cost and capacity.
SSDs are inferior to HDDs.
Capacity
Low weight
SSD contains
no mechanical
parts.
Low noise
100%
80%
60%
40%
20%
0%
Low $/GB
Read
− SSD
− HDD
Operating…
Write
Small size
Low Op. power
MTBF
Source: Inside NAND Flash Memories
8
Capacity
Capacity: HDD and SSD
100%
50%
0%
− SSD
− HDD
• Average capacity of HDDs is about 10 times
larger than that of SSDs.
HDD
Now
Source: Edward Grochowski, Flash Memory Summit, 2012
9
SSD Now
Source: Forward Insights, 2014
Capacity
Capacity: NAND Flash Memory
100%
50%
0%
− SSD
− HDD
• Manufacturers are developing 256Gb 3D NAND
flash memories.
2D NAND
Now
3D NAND
Source: Edward Grochowski, Flash Memory Summit, 2012
10
Capacity
Capacity: NAND Flash Memory
100%
50%
0%
− SSD
− HDD
• 3D NAND technology can increase capacity
without scaling down.
3D NAND
Source: ES Jung, Flash Memory Summit, 2013
11
$/GB: HDD and SSD
100%
50%
0%
Low $/GB
− SSD
− HDD
$/GB
• In terms of price, SSD will be comparable to HDD
in 2017.
Storage SSD: SAS based
Server SSD: SATA based
Mission Critical HDD: SAS based
Business Critical HDD: SATA based
Source: Gartner, Market Trends, 2013
12
$/GB: NAND Flash Memory
100%
50%
0%
Low $/GB
− SSD
− HDD
• Cost reduction of NAND flash memory will
continue at a steep rate via 3D NAND
technology.
2D NAND
3D NAND
Source: Jung H. Yoon, Flash Memory Summit, 2015
13
100%
Performance: Interface
50%
0%
Read
Write
− SSD
− HDD
• Interface speed has increased twice over the past
five years.
Interface
Spec.
2009
SATA
Rev 3.0
(600 MB/s)
2010
2011
2012
2013
2014
2015
SSD
PCIe
Gen 3
(1 GB/s)
Gen 4*
(2 GB/s)
Rev 3.0
(400 MB/s)
ONFI
Rev 4.0
(800 MB/s)
NAND
Toggle
2.0
(400 MB/s)
* Not yet released
14
100%
Performance: SATA SSD
50%
Read
0%
Write
− SSD
− HDD
• Performance of SATA SSDs is higher than that of
HDDs.
SC300 series
of SK hynix
128-KB sequential read
128-KB sequential write
4-KB random read
600
100
500
80
IOPS [K]
MB/s
400
300
200
60
40
100
20
0
0
128 GB
256 GB
512 GB
128 GB
15
4-KB random write
256 GB
512 GB
100%
Performance: PCIe SSD
50%
Read
0%
Write
• Performance of PCIe SSDs is even higher than
that of SATA SSDs.
− SSD
− HDD
PC300 series of
SK hynix
(PCIe 4 lanes)
128-KB sequential write
4-KB random read
2500
250
2000
200
1500
150
IOPS [K]
MB/s
128-KB sequential read
1000
100
500
50
0
0
128 GB
256 GB
512 GB
128 GB
16
4-KB random write
256 GB
512 GB
Power Consumption
100%
50%
0%
• SSDs are more energy-efficient than HDDs
because SSDs contain no mechanical parts.
− SSD
− HDD
Low Op.
power
Source: http://adrianotto.com/2013/01/ssd-power-savings-pays-for-itself/
17
100%
Reliability: MTBF
50%
− SSD
− HDD
0%
MTBF
• MTBF is the predicted elapsed time between
inherent failures of a system during operation.
− Specification: 1.5M ~ 2M hours (= 170 ~ 230 years)
• No mechanical parts guarantees more reliable
operation.
Failure
Repair
Failure
Time
MTTR 1)
MTTF 2)
MTBF 3)
1) Mean Time To Repair
2) Mean Time To Failure
3) Mean Time Between Failure
18
100%
Form Factor
50%
0%
− SSD
− HDD
Small size
• 2.5” form factor of SSDs is the same size.
• mSATA and M.2 form factors are much smaller.
Next generation form factor
means M.2 form factor
Source: Edward Grochowski, Flash Memory Summit, 2012
19
100%
Form Factor (Cont.)
50%
− SSD
− HDD
0%
Small size
22 mm
22 mm
22 mm
22 mm
20
22 mm
BGA
SSD
20 mm
30 mm
42 mm
60 mm
80 mm
110 mm
• There are several M.2 form factors for
various capacities and applications.
• BGA SSD is the next step of M.2 SSD.
16 mm
Endurance: NAND Flash Memory
• NAND flash memory is paper-like storage.
− Write after erase
− Limited program/erase cycles
Source: Edward Grochowski, Flash Memory Summit, 2012
21
Source: http://blog.daum.net/daewook/16085169
Endurance: SSD
• TBW (terabyte written) is the total amount of
data written to the SSD over lifespan
• Specification is 75 ~ 150 TBW
− Guarantees 20 years under
the condition of writing 10 GB
per day
•
A typical office user only writes
about 7 GB on average per day
and the number of people who
write over 20 GB is only a few
percent.
Source: http://www.anandtech.com/show/8602/the-state-of-sandisk
22
Endurance: 3D NAND Flash Memory
• Vertical stacking allows to store large number of
electrons per cell independent of scaling, which
makes it possible to keep endurance.
Source: Michael Abraham, Flash Memory Summit, 2015
23
NVMe (NVM Express)
• NVMe optimized for SSDs is a register-level
interface that allows host software to
communicate with a nonvolatile memory system.
Processor-memory complex
Processor-memory complex
PCIe
AHCI
AHCI HBA
SATA
SATA
SATA
SATA
SATA
HDD
HDD
HDD
HDD
PCIe
NVMe
PCIe
AHCI
NVMe
NVMe
NVMe
PCIe
SSD
PCIe
SSD
PCIe
SSD
NVMe

PCIe
SSD
NVMe: Performance
• NVMe can reduce latency overhead by more
than 50%.
− SCSI/SAS: 6.0 us
− NVMe: 2.8 us
Source: Amber Huffman, Flash Memory Summit, 2012
NVMe: Data Set Management
• If host provides useful information to SSDs, SSDs
can optimize read/write operations and improve
both performance and reliability.
Attribute
Description
Command access size
Number of logical blocks expected to be transferred in a single
read or write command
Write prepare
LBA range expected to be written in the near future
Sequential read/write
Dataset optimized for sequential read/write access
Access latency
Expected read/write latency (shorter, typical, longer)
Access frequency
Temporal locality of read/write reference
26
NVMe: Host Memory Buffer
• SSD suppliers started to launch DRAM-less SSD
for low cost application.
• To minimize performance drop, NVMe provides
host memory buffer feature which allows the SSD
controller to utilize an assigned portion of host
memory exclusively.
Host system
DRAM
For
SSD
DRAM
SSD
controller
SSD
NAND flash
memory array
Summary
• SSD technology is evolving rapidly and will hold
up the “Tech Planet” like Atlas.
− 3D NAND technology for low
cost and high reliability
− NVMe-PCIe technology for
high performance.
− M.2 form factors for small size
Source: Atlas and the earth, wikia
28