Upload
others
View
7
Download
0
Embed Size (px)
Citation preview
How vibration affects Hadoop storage performance
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
Gus Malek-MadaniFounder, CTO/CEO
And
Amir YoussefiCo-Founder, VP of Software and Electronics
Green Platform Corporation2455 Old Middlefield Way #SMountain View CA [email protected]
(650)967-4628
Gus Malek-Madani
� 25 years experience in vibration mitigation and carbon fiber design
� Multiple patents related to carbon fiber
▪ Founder and CEO of three companies1. Green Platform Corp: Vibration management for disk storage
2. Composite Products: Improved high-end audio/video performance
3. Composite Rotor: Centrifuges and Rotors for Biotech
Presenters
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
3. Composite Rotor: Centrifuges and Rotors for Biotech
Amir Youssefi
▪ Distributed Systems Architect with expertise in Big Data, Hadoop/MapReduce, Cloud Computing and Data Warehousing at Fortune/Inc 500 companies such as Yahoo, Apple and Plateau
▪ Led software development team for Grid Management System (GMS) operating on Yahoo Hadoop Clusters
2
Providing enterprises with
superior price/performance
Hadoop platforms for Big Data
Green Platform Corporation
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
Hadoop platforms for Big Data
3
• Reduces signal/noise ratio• Reduces performance• Shortens product life
Why worry about vibration?
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
• Adds inefficiency • Hard to resolve
Vibration increases cost
4
Vibration Primer
MX" + CX' + KX = F(t)
Mass
Vibration
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.5
Acceleration Damping
Velocity
StiffnessDisplacement
Vibration
Forces
C and K of materials are frequency dependent
Vibration harmonics
CoolingPower
DistributionDisk Drives
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
Increasing
System DensityUPS & Floor
Fans
Unique and Distinctive
6
Storage is the Bottleneck
Wastes hardwareOrder of magnitude
slower
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
Wastes spaceWastes energy
7
• Very strong, stiff and light
• Excellent damper of vibration
• Different properties in different directions
• Product can be highly customized
What’s so cool about Carbon Fiber?
A 6 µm diameter
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
• Hundreds of variations
• Wide ranges of price/performance
• Product performance depends on
fabrication method and quality
A 6 µm diameter
carbon filament
(running from
bottom left to top
right) compared to
a human hair.
8
Novel Technology
• Patented carbon fiber solution
• Dissipates vibration passively up
to 1000X
• Frictionless implementation AVR-1000 Rack
US Patent No.
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
• Frictionless implementation
• Proven “Real World” results
yielding 40%+ improvement in
performance
9
US Patent No. 8,240,490
Hadoop Primer
Apache Hadoop is an open-source software for reliable, scalable, distributed computing.
Hadoop Distributed File System (HDFS): A distributed file system that provides high-throughput access to application data.
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.10
Hadoop Primer
How does it look like in practice?A single Anti Vibration Rack with 5 node Hadoop Cluster
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.11
Sample Yahoo! Hadoop Cluster
Vibration on a commodity Node
• White box commodity node
• 12 2TB 3.5” 7,200 rpm
desktop HDD
• 12 core (24 threads) CPU
• Typical data center vibration
• IOzone benchmarking tool
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
• IOzone benchmarking tool
12
Base
Vibration
Write
Mb/s
Vibration
penalty
Re-write
Mb/s
Vibration
penalty
.25 grms
300-400 Hz
65 64
No 130 50% 119 47%
Vibration on a Hadoop cluster
• A 4-Node Hadoop cluster
• CDH3u3 (hadoop 0.20.2)
• DFSio
Rack Base Max AVR Average AVR effect on
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.13
Rack Base
vibration
Max
vibration on
server g rms
AVR
vibration
drop
Average
Read IO
rate MB/s
AVR effect on
MB/s
Metal Random .25
grms 300-
400 Hz
0.34 51
AVR Random .25
grms 300-
400 Hz
0.16 53% 67 31%Metal rack on
Shake Table
Terasort test under vibration
• 1 TB dataset
• Default 64 MB block size
• Only one job running on cluster. It
gets better with concurrent jobs.
Rack Base Max AVR Reduce AVR
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.14
Rack Base
vibration
Max
vibration
on server g
AVR
vibration
drop
Reduce
time -
min.
AVR
benefit
Metal Random
.25 grms
300-400 Hz
0.34 37
AVR Random
.25 grms
300-400 Hz
0.16 53% 30 23%
Anti-Vibration Rack
On shake table
In Big Data, is slow OK?
• Population disease tracking and control
• Financial fraud detection
• Portfolio analysis
• Law enforcement and crime prevention/response
Defense
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
• Defense
• Cyber security
• Health care delivery
• Pharma research
• Operational planning and strategic decision-making
• Insurance claims and outcomes in health care
15
• Existing approaches to vibration mitigation are no
longer good enough
• Unstructured data relies on disk storage
• Commodity cluster architectures are very susceptible to
vibration
Challenges for the Storage Industry
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
vibration
• Customers are becoming educated about vibration and
demanding effective vibration management
16
• Commodity hardware slowed by vibration
• MB/s is dropped under Hadoop
• Reduce times longer under Hadoop
• Green Platform rack restores hardware performance
• MB/s: 31% faster
Summary
2012 Storage Developer Conference. © Insert Your Company Name. All Rights Reserved.
• Reduce: 23% less time
• Big Data applications are disk drive intense
• Shared Infrastructure is very dense
• Needless hardware sprawl driven by vibration
• Low risk remedy has been tested
17