A Shift in Data Storage Todd Dinkelman Micron Technology, Inc.
A Shift in Data Storage Todd Dinkelman Micron Technology, Inc.
A Shift in Data Storage Todd Dinkelman Micron Technology, Inc.
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
SSDs<br />
A <strong>Shift</strong> <strong>in</strong> <strong>Data</strong> <strong>Storage</strong><br />
<strong>Todd</strong> D<strong>in</strong>kelman<br />
<strong>Micron</strong> <strong>Technology</strong>, <strong>Inc</strong>.<br />
Santa Clara, CA USA<br />
August 2008 1
Agenda<br />
� SSD Def<strong>in</strong>ed<br />
� Quick Comparison<br />
� Why MLC<br />
� Reliability/Endurance<br />
� Summary<br />
Santa Clara, CA USA<br />
August 2008 2
What is Be<strong>in</strong>g Called an SSD?<br />
� USB <strong>in</strong>terface embedded solution<br />
� Module-based<br />
� IDE <strong>in</strong>terface card w/o enclosure<br />
� 1.8-<strong>in</strong>ch and 2.5-<strong>in</strong>ch, 32GB and 64GB SSD<br />
for notebook and performance comput<strong>in</strong>g<br />
� Module/card with custom form factor, density,<br />
and <strong>in</strong>terface<br />
Santa Clara, CA USA<br />
August 2008 3
Sp<strong>in</strong>dle<br />
Slider (and head)<br />
Actuator<br />
arm<br />
Actuator<br />
SATA <strong>in</strong>terface<br />
connector<br />
Source: Panther Products<br />
Base cast<strong>in</strong>g<br />
The Complexity of HDDs<br />
Actuator axis<br />
Cover mount<strong>in</strong>g holes<br />
Power connector<br />
Platters<br />
Flex circuit<br />
(attaches heads<br />
to logic board)<br />
� HDD Advantages<br />
• Density<br />
• Price/GB<br />
Santa Clara, CA USA<br />
August 2008 4
SATA <strong>in</strong>terface<br />
connector<br />
The Simplicity of SSDs<br />
Controller<br />
Pr<strong>in</strong>ted Circuit<br />
NAND<br />
� SSD Advantages<br />
• Performance<br />
• Size<br />
• Weight<br />
• Ruggedness<br />
• Temperature<br />
Range<br />
• Power<br />
Santa Clara, CA USA<br />
August 2008 5
Capacity<br />
Performance<br />
Reliability<br />
Endurance<br />
Power<br />
Size<br />
Weight<br />
Shock<br />
Temperature<br />
Cost per bit<br />
SSDs vs. HDDs<br />
SSDs<br />
�<br />
�<br />
�<br />
�<br />
�<br />
�<br />
�<br />
�<br />
HDDs<br />
Santa Clara, CA USA<br />
August 2008 6<br />
�<br />
�<br />
�<br />
� Due to recent advances<br />
<strong>in</strong> NAND lithography,<br />
SSD densities have<br />
reached capacities for<br />
mass market appeal<br />
� SSDs offer many<br />
features that lead to<br />
improved user<br />
experiences<br />
� Early shortcom<strong>in</strong>gs<br />
regard<strong>in</strong>g reliability and<br />
endurance are be<strong>in</strong>g<br />
overcome
SSDs <strong>in</strong> Comput<strong>in</strong>g<br />
Relative<br />
Latency<br />
1,200<br />
L2 Cache<br />
DRAM<br />
1,400<br />
Santa Clara, CA USA<br />
Cost/bit data as of November 2007<br />
August 2008 7<br />
1<br />
2.5<br />
CPU<br />
L1 Cache<br />
Relative<br />
Cost/Bit<br />
1,800<br />
NAND 25,000<br />
SSD 3<br />
25,000,000<br />
HDD<br />
10<br />
NAND Flash closes the latency gap<br />
1
NAND <strong>in</strong> Computer Architecture<br />
DRAM<br />
DRAM<br />
USB<br />
Flash Drive<br />
DDR3<br />
USB<br />
CPU<br />
Memory<br />
& Bus<br />
Control<br />
L<strong>in</strong>k<br />
Peripheral<br />
& Bus<br />
Control<br />
SATA<br />
Hybrid<br />
HDD SSD<br />
PCIe Graphics<br />
PCIe<br />
NVMHCI<br />
Flash<br />
Santa Clara, CA USA<br />
August 2008 8
MLC and SLC Differences<br />
� SLC<br />
• S<strong>in</strong>gle-level cell<br />
– One bit per cell<br />
� MLC<br />
• Multi-level cell<br />
– 2 bits per cell – today<br />
– 3 and 4 bits per cell – future<br />
� Endurance<br />
• SLC is typically 10 times better than MLC<br />
� Performance<br />
� Price<br />
• SLC provides ~2X the write performance of MLC<br />
• SLC-based products have better than 2X the<br />
price/GB compared to MLC<br />
Santa Clara, CA USA<br />
August 2008 9
SSD Market Trends<br />
� Improvements <strong>in</strong> controller technology<br />
• Mov<strong>in</strong>g from CompactFlash architectures to true<br />
SSD controllers<br />
� Notebooks and PCs<br />
• Migration to MLC<br />
– Light usage model<br />
– Cost, size, and performance are all important<br />
• Value Proposition<br />
– Better than desktop performance <strong>in</strong> an ultra-light<br />
notebook<br />
Santa Clara, CA USA<br />
August 2008 10
$/GB<br />
SSD Average Price/GB<br />
1.8" HDD<br />
MLC SSD<br />
2004 2006 2008 2010 2012 2014<br />
Santa Clara, CA USA<br />
August 2008 11
Reliability<br />
Endurance<br />
NAND Reliability and Endurance<br />
Effect<br />
Program<br />
Disturb<br />
Read<br />
Disturb<br />
<strong>Data</strong><br />
Retention<br />
Endurance/<br />
Cycl<strong>in</strong>g<br />
Description<br />
Cells not be<strong>in</strong>g<br />
programmed<br />
receive charge<br />
via elevated<br />
voltage stress<br />
Cells not be<strong>in</strong>g<br />
read receive<br />
charge via<br />
elevated voltage<br />
stress<br />
Charge loss over<br />
time<br />
Cycles cause<br />
charge trapped <strong>in</strong><br />
dielectric<br />
Observed As<br />
<strong>Inc</strong>reased read<br />
errors immediately<br />
after programm<strong>in</strong>g<br />
<strong>Inc</strong>reased read<br />
error at high<br />
number of reads<br />
<strong>Inc</strong>reased read<br />
errors with time<br />
Failed<br />
program/erase<br />
status<br />
Management<br />
ECC and Block<br />
Management<br />
ECC and Block<br />
Management<br />
ECC and Block<br />
Management<br />
Retire Block<br />
Santa Clara, CA USA<br />
August 2008 12
NAND Error Rate<br />
� Bit Error Rate<br />
– Fail<strong>in</strong>g bits corrected with appropriate levels of ECC<br />
– Correctable bit errors do not result <strong>in</strong> data loss<br />
� Raw Bit Error Rate (RBER): Bit error rate<br />
prior to ECC<br />
� Uncorrectable Bit Error Rate (UBER): Bit<br />
error rate after ECC<br />
� UBER is projected us<strong>in</strong>g the measured RBER<br />
and specific level of ECC<br />
Santa Clara, CA USA<br />
August 2008 13
SSD Reliability and Endurance<br />
� SSD reliability has two parts<br />
• MTBF<br />
– Measure of time between failures due to manufactur<strong>in</strong>g or<br />
component defects<br />
– 2 million hours is typical for <strong>Micron</strong> SSDs<br />
• Endurance<br />
– SSDs all wear out due to data writes<br />
– Indication of drive life based on a usage condition<br />
– <strong>Micron</strong> SSDs are specified to last 5 years under predef<strong>in</strong>ed<br />
usage conditions<br />
– Usage conditions vary for consumer and performance products<br />
Santa Clara, CA USA<br />
August 2008 14
Endurance Factors<br />
� Wear-level<strong>in</strong>g efficiency<br />
� Write amplification<br />
� NAND cycles<br />
� SSD densities<br />
Santa Clara, CA USA<br />
August 2008 15
Wear-Level<strong>in</strong>g<br />
� Dynamic wear-level<strong>in</strong>g uses the available<br />
free space and the <strong>in</strong>com<strong>in</strong>g data to equally<br />
wear each of the physical blocks<br />
• Static data is not <strong>in</strong>cluded <strong>in</strong> the available pool of<br />
wear-level<strong>in</strong>g blocks, leav<strong>in</strong>g a portion of the drive<br />
with no wear<br />
� Static wear-level<strong>in</strong>g considers all physical<br />
blocks <strong>in</strong> the SSD, regardless of content, and<br />
ma<strong>in</strong>ta<strong>in</strong>s an even level of wear across the<br />
entire drive<br />
Santa Clara, CA USA<br />
August 2008 16
Wear-Level<strong>in</strong>g Example<br />
� MLC devices can typically support 10,000 cycles per<br />
block<br />
� If you erased and reprogrammed one block every<br />
second, you would exceed the 10,000 cycl<strong>in</strong>g limit <strong>in</strong><br />
just 3 hours!<br />
60 x 60 x 3 = 10,080<br />
� Rather than cycl<strong>in</strong>g the same block, wear-level<strong>in</strong>g<br />
<strong>in</strong>volves distribut<strong>in</strong>g the number of blocks that are<br />
cycled<br />
Santa Clara, CA USA<br />
August 2008 17
Wear-Level<strong>in</strong>g Example (cont<strong>in</strong>ued)<br />
� An 8GB MLC-based SSD conta<strong>in</strong>s 32,768 <strong>in</strong>dependent<br />
blocks (each block is 256KB of data)<br />
� If we took the previous example and distributed the cycles<br />
over all 32,768 blocks, each block would have been<br />
programmed once after 9 hours<br />
� If you provided perfect wear-level<strong>in</strong>g on an 8GB drive, you<br />
could erase and program a block every second, every day<br />
for over 10 years!<br />
10,000 X 32,768 327,680,000<br />
= = 3,792 days = 10.38 years<br />
60 X 60 X 24 86,400<br />
Santa Clara, CA USA<br />
August 2008 18
Write Amplification<br />
� Additional write overhead due to garbage collection<br />
and wear-level<strong>in</strong>g<br />
• Write Amplification = Flash Writes / Host Writes<br />
� Influence on host data patterns<br />
• Large number of small random transfers <strong>in</strong>creases write<br />
amplification<br />
� Influenced by number of spare blocks and static data<br />
• Keep<strong>in</strong>g a percentage of the drive as spares will bound the<br />
write amplification<br />
Santa Clara, CA USA<br />
August 2008 19
1800<br />
1600<br />
1400<br />
1200<br />
1000<br />
800<br />
600<br />
400<br />
200<br />
0<br />
Typical PC File Transfers<br />
Less than<br />
4KB<br />
4KB to<br />
16KB<br />
Avg. Disk Bytes Per Read/Write<br />
16KB to<br />
32KB<br />
32KB to<br />
64KB<br />
Transfer Size<br />
64KB to<br />
128KB<br />
128KB to<br />
256KB<br />
Greater<br />
than<br />
256KB<br />
Reads<br />
Writes<br />
Santa Clara, CA USA<br />
August 2008 20
SSD Endurance Calculation<br />
� Write amplification and wear-level<strong>in</strong>g<br />
efficiency must be accounted for when<br />
calculat<strong>in</strong>g SSD lifetime<br />
Life <strong>in</strong> Years =<br />
NAND Cycles * SSD Capacity<br />
Amplification Factor * GB/Year<br />
Santa Clara, CA USA<br />
August 2008 21
SSD Lifetime<br />
� Given a capacity po<strong>in</strong>t, NAND endurance, and writes-per-day, the<br />
upper limit on SSD lifetime is def<strong>in</strong>ed<br />
• Example: A 64GB SSD, 10K PROGRAM/ERASE cycles and write duty of<br />
350 GB/day will wear out <strong>in</strong> 5 years<br />
� Uneven wear with<strong>in</strong> a NAND Flash device may cause ‘hot spots’<br />
• Wear-level<strong>in</strong>g of NAND is necessary<br />
� Improper NAND data management may cause lifetime to be<br />
significantly less than the limit<br />
• A Flash ERASE operation must occur before a page can be rewritten<br />
• To improve overall throughput, erase granularity is much larger than write<br />
granularity<br />
• LBA traffic pattern can cause lots of extra data movement<br />
Santa Clara, CA USA<br />
August 2008 22
Times per Day (Count)<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
SSD Endurance<br />
Average Number Of Times Per Day The Drive Space Gets Written<br />
0 12 24 36 48 60 72 84 96<br />
Required Lifetime (Months)<br />
33% WRITEs<br />
2K random IOPs<br />
64GB drive<br />
(55GB user-accessible<br />
20% reserved spares<br />
Santa Clara, CA USA<br />
August 2008 23
Time to Failure<br />
SSD Capacity Effect on Endurance<br />
Years for SSD to Wear Out<br />
32 64 128 256 512<br />
Drive Capacity (GB)<br />
� SSD life will double with every doubl<strong>in</strong>g of capacity<br />
Santa Clara, CA USA<br />
August 2008 24
Summary<br />
� SSDs br<strong>in</strong>g many advantages to storage<br />
• Performance<br />
• Power<br />
• Reliability<br />
• Environmental ruggedness<br />
� Different products for different markets<br />
• MLC-based SSDs for PCs<br />
� Reliability has two aspects<br />
• Endurance and MTBF<br />
Santa Clara, CA USA<br />
August 2008 25
Thank You<br />
Santa Clara, CA USA<br />
August 2008 26