how ssds fit in different data center applications
TRANSCRIPT
How SSDs Fit in Different Data Center Applications
Tahmid Rahman Senior Technical Marketing Engineer
NVM Solutions Group
Flash Memory Summit 2012Santa Clara, CA 1
Tuesday, August 21, 12
Agenda
§ SSD market momentum and drivers§ Placement in server/storage applications§ Application specific requirements and
workload characteristics § Proof points with SSDs in transaction
processing, IT, virtualization§ Call to action
Tuesday, August 21, 12
2008 2009 2010 2011 2012 2013 2014 2015
165664.3000
122760.7000
82984.8500
49616.4600
Mill
ions
We are here
Strong SSD Momentum in the Market
Source: Gartner Q1’12
Worldwide SSD Unit Sales1
Tuesday, August 21, 12
2008 2009 2010 2011 2012 2013 2014 2015
165664.3000
122760.7000
82984.8500
49616.4600
Mill
ions
We are here
Strong SSD Momentum in the Market
Source: Gartner Q1’12
Worldwide SSD Unit Sales1
Tuesday, August 21, 12
2008 2009 2010 2011 2012 2013 2014 2015
165664.3000
122760.7000
82984.8500
49616.4600
Mill
ions
We are here
Strong SSD Momentum in the Market
Source: Gartner Q1’12
Worldwide SSD Unit Sales1
Tuesday, August 21, 12
2008 2009 2010 2011 2012 2013 2014 2015
165664.3000
122760.7000
82984.8500
49616.4600
Mill
ions
We are here
Strong SSD Momentum in the Market
Source: Gartner Q1’12
Worldwide SSD Unit Sales1
Tuesday, August 21, 12
Drivers Behind Data Center Storage
Flash Memory Summit 2012Santa Clara, CA 4
§ Architectural Changes – § Big data§ Cloud§ Software innovation for caching, tiering
§ Server Side Innovations-§ De-duplication, compression§ Thin-provisioning§ Virtualization
§ Interface transitions§ SATA/SAS to PCIe§ AHCI based to NVMe
§ SSD endurance and performance grades§ Endurance classes – high, medium, standard§ Optimization for access – read intensive, write intensive,
mixed workload§ Different “out of the factory” spare area level
Tuesday, August 21, 12
Usage Applications Compute(Servers)
External Storage
Cache
(Low, Deterministic Latency, $/IOP
IPDC Web 2.0Volume HPCVirtualized SAN/NASFSI
Persistent cache:(Block Cache, User space buffer cache, NFS v4 cache)
Persistent Cache:(e.g. OS block, metadata, de-dupe, etcPerformance
($/IOP/GB) IPDC web2.0Volume HPCOLTPOLAPCDNVOD
Hot Application Data(Web, Database, Email, Search, Videos, IPDC etc)
Hot Application Data(Database, Email, etc)
Capacity
($/TB, Watt/TB) Data WarehouseBackupsArchives
Luke-warm Application Data(Web, Email, Videos, etc)
Cold/Luke-warm
Boot ($/GB) All Server Applications
Local boot data(Operating System, Hypervisor, SWAP, VM, Application Image)
Local boot Data: (SAN/NAS image)
SSD Placement in Server/Storage Application
Tuesday, August 21, 12
Highest Requirements for Data Center SSDs
§ Data Integrity§ True End to End data protection§ Power Loss Protection§ Power loss cap self test§ Protection of internal memory with ECC and parity
§ Predictable Performance§ IOPS variation needs to be within a narrow range§ Latency outliers should be within a max value
§ High Endurance Requirement§ Two primary endurance evolving
§ Standard endurance 0.1-1 DWD § High endurance 10 DWD
Tuesday, August 21, 12
Data Center Application Workload Characteristics
Applications Transfer Size % Random % Read Write.
EnduranceQuality of
service
Media Streaming 64KB Low High Med Med
Web-server Logging 8KB Low Low Med Med
Search Engine 4KB/8KB/16KB High High Low High
Video-On-Demand 128KB High High Low High
Caching 512KB High High Low Med
Decision Support 64KB High High Low High
Content Delivery Network 16KB/32KB High Mixed High High
Database OLTP(On Line
Transaction Processing)
4KB/8KB High Mixed High High
7
Sequential
Random Read
Mixed Random
Source: Industry Standard Benchmarks and Customer Engagement DataPatterns will vary for unique customers
Tuesday, August 21, 12
SSDs For Virtual Storage in the Cloud
Challenges§ Reversed server to data store ratio ( multiple VMs running on single array)§ Adding storage and cache is cost prohibitive
Solutions§High Performance SSD 3x8 RAID5 meeting multiple VM random IOPS of ~100K w/ SW SAN solution
Impact§Expanded performance at a lower cost >75% TCO reduction
– 450 15K RPM HDDs vs. 24 Intel 710 SSDs– IT professional would spend $43K instead of $200K+
Tuesday, August 21, 12
SSDs for Transaction Processing
§ TPoX* (Transactional Processing over XML*) is an application-based benchmark that mimics a storage-bound online transaction procession over XML data for brokerage
§ Intel® SSD 910 Series, reveals a replacement ratio of 1 to 180 with Standard Magnetic Drive Solutions
§ 1 TB database can be compressed in one single PCIe card and meet the performance of 180 magnetic storage 15K RPM SAS drives
PCIe SSD Magnetic Storage
15K RPM SAS
Drives
Number of drives 1 180TPS score (steady-state run) 13,516 13,742Latency (msec) 5.93 7.15Drive Cost 3,859 59,000Storage Subsystem Cost X 14,000
Configuration: Intel® Xeon® Processor X5680 (3.33 GHz, 6.40 GT/s Intel® QPI) platforms with Intel® 7500 Chipset, 72GB (18x4GB), 800MHz DDR3 memory, SUSE SLES 11 SP1 operating system, DB2 9.7, and TPoX 2.0 using “M” factor scale (1 TB data size). Hitachi* HUS151P1 CLAR146 146GB SAS 15K RPM drives.
Server:Exercising Application Load
4 x 4Gb/s
“Fiber Channel”
Database Storage:1 PCIe SSD
Server:Exercising Application Load
PCIe SSD Based SolutionHDD Based Solution
http://www.intel.com/content/dam/www/public/us/en/documents/technology-briefs/ssd-910-tpox-brief.pdf
Tuesday, August 21, 12
Transaction Processing And Importance of Latency QoS
§ Transaction processing requires dense IO (Higher IOPS/GB)
§ Systems tune to have no “storage bottleneck”
§ No Mercy for latency outliers and occasional drops of IOPS
0%
25.00%
50.00%
75.00%
100.00%
0.9999 0.99999 0.999999
99.12%94.71%
72.30%
% T
rans
actio
ns p
er u
nit t
ime
% of workload completed within 90 msec
0%
25.00%
50.00%
75.00%
100.00%
10 20 30 40 50 90
72.30%
84.53%88.51%92.72%96.77%98.99%
% T
rans
actio
ns p
er u
nit t
ime
99.99 percentile latency (msec)
Acknowledgement: Terry Pricket, Jeff Smit, Intel SSG Lab
Tuesday, August 21, 12
Transaction Processing And Importance of Latency QoS
§ Transaction processing requires dense IO (Higher IOPS/GB)
§ Systems tune to have no “storage bottleneck”
§ No Mercy for latency outliers and occasional drops of IOPS
0%
25.00%
50.00%
75.00%
100.00%
0.9999 0.99999 0.999999
99.12%94.71%
72.30%
% T
rans
actio
ns p
er u
nit t
ime
% of workload completed within 90 msec
0%
25.00%
50.00%
75.00%
100.00%
10 20 30 40 50 90
72.30%
84.53%88.51%92.72%96.77%98.99%
% T
rans
actio
ns p
er u
nit t
ime
99.99 percentile latency (msec) Gain transactions by 27% by moving to less frequent occurrence of outlier event (99.99% to
99.9999%)
Acknowledgement: Terry Pricket, Jeff Smit, Intel SSG Lab
Tuesday, August 21, 12
Transaction Processing And Importance of Latency QoS
§ Transaction processing requires dense IO (Higher IOPS/GB)
§ Systems tune to have no “storage bottleneck”
§ No Mercy for latency outliers and occasional drops of IOPS
0%
25.00%
50.00%
75.00%
100.00%
0.9999 0.99999 0.999999
99.12%94.71%
72.30%
% T
rans
actio
ns p
er u
nit t
ime
% of workload completed within 90 msec
0%
25.00%
50.00%
75.00%
100.00%
10 20 30 40 50 90
72.30%
84.53%88.51%92.72%96.77%98.99%
% T
rans
actio
ns p
er u
nit t
ime
99.99 percentile latency (msec) Gain transactions by 27% by moving to less frequent occurrence of outlier event (99.99% to
99.9999%)
Acknowledgement: Terry Pricket, Jeff Smit, Intel SSG Lab
Tuesday, August 21, 12
Transaction Processing And Importance of Latency QoS
§ Transaction processing requires dense IO (Higher IOPS/GB)
§ Systems tune to have no “storage bottleneck”
§ No Mercy for latency outliers and occasional drops of IOPS
0%
25.00%
50.00%
75.00%
100.00%
0.9999 0.99999 0.999999
99.12%94.71%
72.30%
% T
rans
actio
ns p
er u
nit t
ime
% of workload completed within 90 msec
0%
25.00%
50.00%
75.00%
100.00%
10 20 30 40 50 90
72.30%
84.53%88.51%92.72%96.77%98.99%
% T
rans
actio
ns p
er u
nit t
ime
99.99 percentile latency (msec) Gain transactions by 27% by moving to less frequent occurrence of outlier event (99.99% to
99.9999%)
Gain transactions by 24% by moving the latency from 90 msec to 20 msec
Acknowledgement: Terry Pricket, Jeff Smit, Intel SSG Lab
Tuesday, August 21, 12
SSDs for IT Management Services
• Automatic Updates for IT security patches
• Managing Design Simulation database
• Swap operation for over-flow memory
• Benchmarking and proof points
Tuesday, August 21, 12
Enterprise Patching and Security Compliance Performance Comparison
Flash Memory Summit 2012Santa Clara, CA 12
0
75
150
225
300
1 90 179
268
357
446
535
624
713
802
891
980
1069
1158
1247
1336
1425
1514
1603
1692
1781
1870
1959
2048
2137
2226
2315
2404
2493
2582
2671
2760
2849
2938
3027
3116
3205
3294
3383
3472
3561
3650
3739
3828
3917
4006
4095
4184
4273
4362
With 15K RPM Drive
Que
ue D
epth
s
0
75
150
225
300
1 90 179
268
357
446
535
624
713
802
891
980
1069
1158
1247
1336
1425
1514
1603
1692
1781
1870
1959
2048
2137
2226
2315
2404
2493
2582
2671
2760
2849
2938
3027
3116
3205
3294
3383
3472
3561
3650
3739
3828
3917
4006
4095
4184
4273
4362
With SSDs deployed
Que
ue D
epth
s
Acknowledgement: Christian Black, Intel IT Architect
QD gets built up easily and the IO demand cannot be met
QD is never too deep, IO saturation maintained without it
Tuesday, August 21, 12
Are All SSDs Ready for Transaction Processing?
Tuesday, August 21, 12
Are All SSDs Ready for Transaction Processing?
“Zero” IOPS!
Tuesday, August 21, 12
Are All SSDs Ready for Transaction Processing?
“Zero” IOPS!
ß RAID Array stalls and timeouts
ß Higher drive counts to meet IO needs
Tuesday, August 21, 12
Are All SSDs Ready for Transaction Processing?
“Zero” IOPS!
1 sec max latency!
ß RAID Array stalls and timeouts
ß Higher drive counts to meet IO needs
SSD B
Tuesday, August 21, 12
Are All SSDs Ready for Transaction Processing?
“Zero” IOPS!
1 sec max latency!
ß RAID Array stalls and timeouts
ß Higher drive counts to meet IO needs
Negative SLA impacts à
Catastrophic for certain applications à
SSD B
Tuesday, August 21, 12
Call to Action
§ Ample opportunity for SSD proliferation within data center
§ Innovate around applications needs§ Use faster interface and technology to unleash NAND
backend bandwidth
Tuesday, August 21, 12