15 - Sun - Ouwens
-
Upload
lara-salinas -
Category
Documents
-
view
220 -
download
0
Transcript of 15 - Sun - Ouwens
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 1/31
1
1
BIG DATA
The Petabyte Age:More Isn't Just More —More Is Differentwired magazine Issue 16/.07
Martien OuwensDatacenter Solutions Architect
Sun Microsystems
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 2/31
Sun Confidential: Internal Only 2
Next Generation
Big Data, The Petabyte Age:
Relational
The biggest challenge of the PetabyteAge won't be storing all that data, it'll be
figuring out how to make sense of itwired magazine Issue 16/.07
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 3/31
Sun Confidential: Internal Only 3
Next Generation
Big Data Key Trends
Relational
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 4/31
Sun Confidential: Internal Only 4
“Hitting walls” in Processor DesignClock frequency
− frequency increases tapering off, in newsemiconductor processes
− high frequencies => power issuesMemory latency (not instruction execution speed)dominating most application timesProcessor designs for high single-threadperformance are becoming highly complex,therefore:
− expense and/or time-to-market suffer− verification increasingly difficult− more complexity => more circuitry => increased
power ... for diminishing performance returns
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 5/31
Sun Confidential: Internal Only 5
Memory BottleneckRelativePerformance
10000
1
1990 199520051980
1000
100
10
1985 2000
Gap
CPU FrequencyDRAM Speeds
C P U - - 2
x E v e r y 2 Y
e a r s
D RA M - - 2 x E v e r y 6 - 8 Y e a r s
Source: Sun World Wide Analyst Conference Feb. 25, 2003
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 6/31
Sun Confidential: Internal Only 6
How We Mask Memory Latency Today
L2 CacheL2 Cache(Left)(Left)
L2 TagsL2 TagsM C U
M C U
S I U S I U
SIUSIU
L2/L3 CntlL2/L3 Cntl
L1 IL1 I--CacheCache L1L1DD--CacheCache
D T L B
D T L B
ITLBITLB
P$P$
W$W$
Core 0Core 0
L2 CacheL2 Cache(Right)(Right)
L3 TagsL3 Tags
L1 IL1 I--CacheCache ITLBITLB
P$P$
W$W$
D T L B
D T L B
• of many different kinds• that only mask memory
latency• and execute no code• accessed one cache
line at a time• but require power and
cooling to all lines allthe time
Core 1Core 1
L1L1DD--CacheCache
Great Big CachesGreat Big Caches
Cache Logic Accounts for About 75% of the Chip AreaCache Logic Accounts for About 75% of the Chip Area
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 7/31
Sun Confidential: Internal Only 7
“Power Wall + Memory Wall + ILP Wall = BrickWall”
* “The Landscape of Parallel Computing Research: A View from Berkeley”, December 2006 (emphasis added)http://www.eecs.berkeley.edu/Pubs/TechRpts/2006/EECS-2006-183.html
““In 2006, performance is a factor of three below the traditionalIn 2006, performance is a factor of three below the traditionaldoubling every 18 monthsdoubling every 18 monthsthat we enjoyed between 1986 and 2002.that we enjoyed between 1986 and 2002.
The doubling of uniprocessor performance may now takeThe doubling of uniprocessor performance may now take5 years5 years ..””**
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 8/31
Sun Confidential: Internal Only 8
Single ThreadedPerformance
Single Threading
Thread
Memory LatencyCompute
Time
C C CM M M
Up to 85% Cycles Spent Waiting for Memory
SPARC US-IV+ 25% 75%
Intel 15% 85%
Hurry Upand
WAIT
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 9/31
Sun Confidential: Internal Only 9
Comparing Modern CPU DesignTechniques
1 GHz
Time
ILP Offers Limited HeadroomTLP Provides Greater Performance Efficiency
CC MM CC MM CC MM CC MM ComputeMemory Latency
CC MM CC MM CC MM CC MM2 GHz
TLPCC MM
CC MMCC MM
CC MM
Time Saved
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 10/31
Sun Confidential: Internal Only 10
Core-1
Core-2
Core-3
Core-4
Core-5
Core-6
Core-7
Core-8Time
Chip Multi-Threading (CMT) Utilization: Up to 85%*
Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1Thread 4Thread 3Thread 2Thread 1
Latency
UltraSPARC-T1 85% 15%
Compute
Chip Multi-threaded
(CMT)Performance
* based on example of UltraSPARC T1
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 11/31
Sun Confidential: Internal Only 11
Enterprise Flash – Relative Perf
1 sec
16,7 min
11,5 dag
31,7 jr
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 12/31
Sun Confidential: Internal Only 12
Where to Store Data?Optimization Trade-Off
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 13/31
Sun Confidential: Internal Only 13
Data Retention Comparison
DRAM SSD HDD
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 14/31
Sun Confidential: Internal Only 14
ZFS Turbo Charges ApplicationsHybrid Storage Pool Data Management on Solaris Systems, OpenStorage
•ZFS automatically:•Writes new data to very fast SSDpool (ZIL)
•Determines data access patternsand stores frequently accesseddata in the L2ARC
•Bundles IO into sequential lazywrites for more efficient use of lowcost mechanical disks
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 15/31
Sun Confidential: Internal Only 15
Unique Capabilities from SunSun Storage 7000 Unified Storage Systems
• Radical simplification> Ready-to-go appliance – installs in minutes> Extensive analytics and drill-down for rapid
management of performance challenges
• Improved latency and performance> Integrated flash drives> Hybrid Storage Pools automates data
placement on the best media
• Compelling value through standards> Industry standard hardware and open source
software vs. proprietary solution
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 16/31
Sun Confidential: Internal Only 16
Sun Oracle Database Machine• Hardware by Sun, software by Oracle – Exadata Version 2
• Enhancements to Exadata Version 1 offering extreme performance> Faster servers, storage, and network interconnects> Flash acceleration> Speedup at Oracle level of 20 times processing IOPS> Extends Exadata to provide acceleration for OLTP
• Database aware storage> Smart scans> Storage indexes – eliminates a good deal of I/Os
• Massively parallel architecture• Dynamic linear scaling
• Offering mission critical availability and protection of data
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 17/31
Sun Confidential: Internal Only 17
• Current database deployments often have bottlenecks limiting the movement of data from
disks to servers> Storage Array internal bottlenecks on processors and Fibre Channel Architecture> Limited Fibre Channel host bus adapters in servers> Under configured and complex SANs
• Pipes between disks and servers are 10x to 100x too slow for data size
What Does Sun Oracle Database MachineAddress?
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 18/31
Sun Confidential: Internal Only 18
• Adding more pipes – Massively parallel architecture
• Make the pipes wider – 5X faster than conventional storage• Ship less data through the pipes – Process data in storage
• Simplify the storage offering – eliminate the need of a SAN
• Allow for Ease of Growth• Provide a Flash Cache for Rapid Response
Sun Oracle Database MachineSolves the Bottleneck by
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 19/31
Sun Confidential: Internal Only 19
HPC Market Trends• It’s a growing market,outpacing growthof many others
> $10B in 2008
> CAGR of 3.1%for 2007 to 2012*
• Clustering is nowthe dominantarchitecture• HPC technologiesnow within reachof all organizations
HPC is Underserved by Moore’s Law
* Source IDC (January 2009)
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 20/31
Sun Confidential: Internal Only 20
Sun Constellation SystemOpen Petascale Architecture
NetworkingCompute Storage Software
Ultra-DenseBlade Platform
• Fastest processors:AMD Opteron,Intel Xeon
• Highest computedensity
• Fastest host
channel adaptor
Ultra-DenseSwitch Solution
• 648 port InfiniBandswitch
• Unrivaled cablesimplification
• Most economicalInfiniBand cost/port
ComprehensiveSoftware Stack
• Integrated developertools
• Integrated Grid Engineinfrastructure
• Provisioning, monitoring,patching
• Simplified inventorymanagement
DEVELOPER TOOLS
GRID ENGINEPROVISIONING
Ultra-DenseStorage Solution
• Most economical andscalable parallel filesystem building block
• Up to 48 TB in 4RU• Up to 2TB of SSD• Direct cabling to
IB switch
Eco-Efficient Building Blocks
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 21/31
Sun Confidential: Internal Only 21
Sun Constellation SystemOpen Petascale Architecture
CompetitorsCablingInfrastructure
CompetitorsCompute
Clusters
ReducedCabling
R a c k s
C a b l i n
g
C o r e S w i t c h e s
L e a f S w i t c h e s
Radical Simplicity, Faster Time to Deployment
R a c k sSun
ConstellationSystemCluster
C a b l i n
g
•Alternative Open Standards Fabric• 300 switching elements• 6912 cables
• 92 racks
• 1 switching element300:1 reduction• 1152 cables6:1 reduction• 74 racks20% smaller footprint
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 22/31
Sun Confidential: Internal Only 22
Massive Low Latency Engine
• 3,936 Sun four-socket,quad-core nodes
• 15,744 AMD Opteron processors• Quad-core, four flops/cycle
(dual pipelines)
Compute Power —579 TFLOPS Aggregate Peak
Memory
• 72 Sun x4500 “Thumper”I/O servers, 24TB each
• 1.7 Petabyte total raw storage• Aggregate bandwidth ~32 GB/sec
Disk Subsystem
InfiniBand Interconnect
• 2 GB/core, 32 GB/node, 125 TB total
• 132 GB/s aggregate bandwidth
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 23/31
Sun Confidential: Internal Only 23
The HPC Data I/O Challenge
• Compute power is scaling rapidly> Larger compute clusters> More sockets per server > More cores per CPU
• Compute advances far outpace those in I/O> Storage latency 12 orders of magnitude higher
than compute
> Compute Side developed Parallel Approachessooner
• Need a better approach to feed compute
> Cluster performance is often limited by data I/O> Need a parallel I/O approach to help balance things
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 25/31
Sun Confidential: Internal Only 25
Unique Capabilities from SunSun Storage Cluster
• High performance and broad scaling> From a few to over 100 GB/second> From dozens to thousands of nodes
• Breakthrough economics, up to 50% savings> Built with standardized hardware and opensource software vs. proprietary products
• Simplifying Lustre for the mainstream> Pre-defined modules speed deployment ofparallel file system, available factory integrated
> Ease of use improvements – patchless clients,
integrated driver stack, LNET self-test, mount-based cluster configuration, many more
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 26/31
Sun Confidential: Internal Only 26
Sandia Red Storm340 TB Storage, 50GB/s I/O throughput,
25,000 clients
German Climate ResearchData Centre (DKRZ)
272 TB, Lustre and Storage Archive Manager,256 nodes with 1024 cores
Framestore“The Tale of Despereaux”
200TB Lustre file system – averaged 1.2GB/s in sustained reads,
peaks of 3GB/s, 5TB data generated per night from clusterof 4,000 cores interfacing with Lustre file system
TACC Ranger 1.73 PB storage, 35GB/s I/O throughput,
3,936 quad-core clients
Lustre in Action
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 27/31
Sun Confidential: Internal Only 27
Next GenerationBig Data Key Trends
Distributed data and processing(Dist. FS, MapReduce, Hadoop...)
RelationalCommoditisation(Open source, COTS)
Proprietary, dedicateddatawarehouse
OLTP is thedatawarehouse
Open source, generalpurpose datawarehouse
Object Store
Distributed FS Federated/Shared
Master/Master
Master/Slave
Unstructured Data StructuredData
Semi-structuredDatabase
ScaleDB, Big Table,SimpleDB hBase
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 28/31
Sun Confidential: Internal Only 28
Big DataMartien OuwensDatacenter Solutions [email protected]
Sun Microsystems Nederland B.V.
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 29/31
Slide 29
Typical BI/DW Components
Load
Transactionalsystems
Line ofbusinesssystems
S
our ce
s
Pentaho Pentaho
Sun Servers Sun Servers
ODS
Repository
Reporting
Data Warehouse
Infobright
Sun Storage
Analytics
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 30/31
Sun Confidential: Internal Only 3030
Smarter architectureLoad data and goNo indexes or partitionsto build and maintain, no tuning
Super-compact data foot- printleverages off-the-shelfhardwareMySQL wrapper connects to BIand ETL tools
Data Packs – data storedin manageably sized, highlycompressed data packs
Knowledge Grid – statistics andmetadata “describing” the super-compressed data which isautomatically updated as data packsare created or updated
Column Orientation
Infobright Optimizer iterates over theKnowledge Grid; only the data packsneeded to resolve the query aredecompressed
Column-OrientationExemplified by InfoBright's approach
8/13/2019 15 - Sun - Ouwens
http://slidepdf.com/reader/full/15-sun-ouwens 31/31
Sun Confidential: Internal Only 31
`employees` records stored in DataPacks of 64k elements each
SELECT COUNT(*) FROM employees
WHERE salary > 50000
AND age < 65
AND job = ‘Shipping’
AND city = ‘TORONTO’;
Using the Knowledge Grid, we verify, which packs are
• irrelevant (no values found that fit SELECT criteria)
• relevant (all values fit)
• suspect (some values fit)
Any row containing 1 or more irrelevant DP’s is ignored.
Since we are looking for COUNT(*) in this case we do notneed decompress relevant packs. COUNT information is keptin the Knowledge Grid in the Data Pack Nodes.
The suspect packs must be decompressed for detailedevaluation.
salary age job city
Relevant packs
Only this pack will bedecompressed
All packsignored
Rows 1 to65,536
65,537 to131,072
131,073 to …
All packs
ignored
All packsignored
InfoBright Query Execution