ssd architectural considerations for data center workloads
TRANSCRIPT
SSD Architectural Considerations for Data
Center Workloads
Paul Truong, Principal Engineer
Data Center Solutions, SanDisk Corp
Santa Clara, CA
August 2015
1
Forward-looking Statements
During our meeting today we will 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 products and their
capabilities, performance and compatibility, cost savings and other benefits to customers.
Actual results may differ materially from those expressed in these forward-looking statements
due to a number of risks and uncertainties, including the 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.
Santa Clara, CA
August 2015
2
Data Center Workloads
Santa Clara, CA
August 2015
3
Cost
Big Data
DB
Real Time
Capacity
Application Filter is a modifier to Value
Big Data Use Cases Where Flash is Displacing Traditional Media… Perspective
Active Archive (Rack) • 15 InfiniFlash ™ System IF100s, 2 Servers
• Capacity 7.5PB (6.8PB Usable) vs. 8TB HDD usable @4+PB
• Bandwidth 225 GB/s vs. HDD @ 30GB/s
• Power 3KW Idle, 8KW All Active vs. HDD @ 9KW
Data Lake Model (Rack) • 8-12 InfiniFlash System IF100s, 8-12 Servers
• Capacity 4-6PB
• Bandwidth up to 180GB/s
• Power 6KW Active
Active Analytics Model (Rack) • 4-8 InfiniFlash System IF100s, 8-24 Servers
• Capacity 2-4PB
• Bandwidth up to 120GB/s
• Power 10KW Very Active
1 GB = 1,000,000,000 bytes. Actual user capacity less. The performance results discussed herein are based on internal testing. Results and performance
may vary according to configurations and systems, including drive capacity, system architecture and applications.
Santa Clara, CA
August 2015
Optimizing Costs for Data Center Workloads
• Capacity & Performance
• Over Provisioning
• Endurance
Santa Clara, CA
August 2015
5
Capacity
• Increase NAND / non-NAND ratio
• 8TB/Controller
• 512TB/Chassis
• Constraints
• Controller Performance (IOPS/TB)
• DRAM Size
Santa Clara, CA
August 2015
6
Right Sizing Performance
Santa Clara, CA
August 2015
7 1 GB = 1,000,000,000 bytes. Actual user capacity less. The performance results discussed herein are based on internal testing. Results and performance
may vary according to configurations and systems, including drive capacity, system architecture and applications.
0
10000
20000
30000
40000
50000
60000
70000
16 32 64 128 256 512
IOP
S/T
B
Application Foot Print (TB)
Application Workload
500K IOPS
700K IOPS
1M IOPS
Right Sizing Capacity
• Typical SSDs @ 80-100K RR IOPs
Santa Clara, CA
August 2015
8
Capacity IOPS/TB
512GB 160K
1TB 80K
2TB 40K
4TB 20K
8TB 10K
Capacity – FTL Implication & Applications
Considerations
• Large logical space for translation table
• 4 bytes entry / 4K logical page
• 1GB DRAM addresses 1TB of logical space
Santa Clara, CA
August 2015
9
Application Solution Option
Real-time Applications Increase DRAM
Big Data Increase granularity of logical page
DB Caching of translation table
1GB
Increase DRAM
Santa Clara, CA
August 2015
10
1GB
1TB
Flash Translation Table
1TB
1TB
1TB
1TB
1TB
1TB
1TB
1TB
DRAM
NAND
1GB
1GB
1GB
1GB
1GB
1GB
4K logical page
Increase Logical Page Size
Santa Clara, CA
August 2015
11
1GB
1GB
1TB
Flash Translation Table
1TB
1TB
1TB
1TB
1TB
1TB
1TB
1TB
DRAM
NAND
16K logical page
Caching of Translation Table
Santa Clara, CA
August 2015
12
1GB
1GB
1TB
Flash Translation Table
1TB
1TB
1TB
1TB
1TB
1TB
1TB
1TB
DRAM
NAND
4K logical page 1GB
1GB
1GB
1GB
1GB
1GB
Endurance at Scale
• Scale:
• 512TB InfiniFlash box
• Bandwidth maxed at 7GB/s
• 90/10 Read/Write workload
• 5 year run, 24/7/365
• Result:
• 100 Petabytes written
• 200 total box writes
Santa Clara, CA
August 2015
13 1 GB = 1,000,000,000 bytes. Actual user capacity less. The performance results discussed herein are based on internal testing. Results and performance
may vary according to configurations and systems, including drive capacity, system architecture and applications.
Lower OP
• Reduced endurance requirement
• Higher sequential write workload
• Enables the use of low 7% OP
Santa Clara, CA
August 2015
14
Read Latencies
• Read latencies during ingest
• High read latencies incur when read access
to die busy with write
• High capacity point reduces probability of
event
Santa Clara, CA
August 2015
15
High Density Chassis
Santa Clara, CA
August 2015
16
• Achieves higher density of Flash in the same chassis than building with
standard SSDs
• A 512TB InfiniFlash System IF100 has 8x higher flash density than today’s densest all-
flash array; and 40% higher density than the most aggressive HDD arrays
• IF100 components are designed to be field replaceable
• Chassis is greater than 1.5 Million Hours MTBF (>99.97% reliability)
• SAS expander boards, fans and drive cards are all hot swappable and built for HA
• Multiple design decisions made to optimize for cost
while providing the required endurance
• Drive card format chosen to maximize “NAND/non-NAND” cost ratio –
• Also allows for 4TB, 8TB, 16TB capacities
• Overprovisioning minimized to 7%
1 GB = 1,000,000,000 bytes. Actual user capacity less. The performance results discussed herein are based on internal testing. Results and performance
may vary according to configurations and systems, including drive capacity, system architecture and applications.
Thank You
Email: [email protected]
LinkedIn: https://www.linkedin.com/pub/paul-truong/4/27/81b
Santa Clara, CA
August 2015
17
1 GB = 1,000,000,000 bytes. Actual user capacity less. The performance results and cost savings discussed herein are based on internal testing. Results and performance may vary
according to configurations and systems, including drive capacity, system architecture and applications.
© 2015 SanDisk Corporation. All rights reserved. SanDisk is a trademark of SanDisk Corporation, registered in the United States and other countries. and others are trademarks of
SanDisk Enterprise IP LLC. InfiniFlash is a trademark of SanDisk Enterprise IP LLC. Other brand names mentioned herein are for identification purposes only and may be the
trademarks of their holder(s).
Back up
Santa Clara, CA
August 2015
18
Performance at Scale
InfiniFlash ™ System IF100 - Purpose built for Big Data
• 64 x 8TB Flash Cards (500TB)
• 780K IOPS 4K Read (aggregate)
• 500+ KIOPS 4K Write (aggregate)
• 7GB/s chassis level bandwidth
• 1/10 AFR of HDD (20HDD failure for every InfiniFlash Card)
• Chassis level idle @ less then 200W, Active @ 400-500W
(Comparative HDD @700W avg)
• Endurance (multi-Exabyte at chassis level)
• 3U
1 GB = 1,000,000,000 bytes. Actual user capacity less. The performance results discussed herein are based on internal testing. Results and performance
may vary according to configurations and systems, including drive capacity, system architecture and applications.
Santa Clara, CA
August 2015
InfiniFlash HW Overview
HSEB
8088
Mini-SAS x4
8088
Mini-SAS x4
8088
Mini-SAS x4
8088
Mini-SAS x4
Expander
PM8004
Gen2 x24
Co
nn
ec
torx4
x4
x4
x4
x8
HSEB
HSEB
8088
Mini-SAS x4
8088
Mini-SAS x4
8088
Mini-SAS x4
8088
Mini-SAS x4
Expander
PM8004
Gen2 x24
Co
nn
ec
torx4
x4
x4
x4
x8
HSEB
FPB Cable
Front Panel Board
BB CableButtons
Status LEDsBB Cable
Bezel
Board
Expander
PM8005
Gen2 x36
Connector
x16
x4
DSEB
Connector
x16
Expander
PM8005
Gen2 x36
Connector
x16
x4
DSEB
Connector
x16
Expander
PM8005
Gen2 x36
Connector
x16
x4
DSEB
Connector
x16
Expander
PM8005
Gen2 x36
Connector
x16
x4
DSEB
Connector
x16
BSSD
Module
BSSD
Module
x32
Port A
BSSD
Module
BSSD
Module
x32
Port A
Port B
Port B
Fa
n C
ab
leF
an
Ca
ble
Fa
n C
ab
leF
an
Ca
ble
Co
nn
ec
torPDB
Co
nn
ec
tor
FPB Cable AC
Power
Module
Co
nn
ec
tor
AC
Power
Module
Fan Cable FIB
Fan
FRU
Fan Cable FIB
Fan
FRU
Fan Cable FIB
Fan
FRU
Fan Cable FIB
Fan
FRU
20
Santa Clara, CA
August 2015