hpc session 7: big data & data center efficiency … data & data center efficiency september...
TRANSCRIPT
© 2014 IBM Corporation
HPC Session 7:Big Data & Data Center Efficiency
September 22, 2014
2© 2014 IBM Corporation
Panel Members
Dave Weber, IBM, x86 Global Financial Segment lead
Rama Karedla, Intel, Performance Architect, HPC/Financial Services
Ed Turkel, HP, Group Manager, HPC Marketing
Nick Papadonis, Oracle, Engineering Consultant
Nick Ciarleglio, Arista, Distinguished Engineer
3© 2014 IBM Corporation
Increase in high profile security breaches
Growth in migration of platforms &applications to cloud service providers
Rise in mobility & Bring Your Own Device(BYOD) increasing data growth and risk
Real-time analysis of huge volumes of data
Enterprise customers assessing ODMs
1 Data from IBM Annual Survey of IT and line-of-business leaders for 20132 Citrix Top Enterprise Mobility Trends, http://blogs.citrix.com/2014/05/06/top-enterprise-mobility-trends/, May 20143 IDC Server Virtualization & Cloud, February 2013
Data center trends
30BConnected devices
by 20203
90%Data on planet created
in last 2 years1
#1Security concern of IT
decision makers & CIOs1
75%IT spend on new cloud
projects by 20172
Security
Cloud
Big Data
Mobility
4© 2014 IBM Corporation
NEW System x3650 M5High-performing & versatile 2U2S rack server for
analytics & cloud
Virtual Desktop
SAP / BusinessAnalytics
DataManagement
Cloud /Virtualization
Big Data
NEW System x3550 M5Compact powerful 1U2S rack server
NEW NeXtScale nx360 M5 ServerVersatile, ½ wide 1U2S server for HPC, cloud, and
analytics
NEW NeXtScale System withWater Cool Technology
Efficient, full-wide, dual-node water cooledserver for HPC
NEW x3500 M5 (1Q15)2S Tower or Rack for
business critical workloads
#1 ReliabilityBuilt-in
Rack & Tower Dense Blade
#1 SecurityBuilt-in
EfficiencyBuilt-in
NEW Flex x240 M5No compromise blade serverfor cloud, virtualization andbusiness applications
New System x portfolio for solutions & business productivity
50%1
More Cores& Cache
2X2
MemoryCapacity
131%3
Faster JavaPerformance
59%4
Faster DBPerformance
39%5
GreaterComputationalPerformance
61%6
GreaterVirtualizationPerformance
50%8
IncreasedMemory
Bandwidth
50%9
Memory PowerSavings
5© 2014 IBM Corporation
Cloud Virtualization
Virtual Desktop
Business Processing
Big Data
Virtualized Storage
Business Continuity
Partnerships with industry leading Independent Software Vendors
6© 2014 IBM Corporation
IBM / Lenovo Announcement Highlights
Lenovo plans to acquire IBM’s x86 server portfolio and related resources and operationsincluding
– System x, BladeCenter and Flex System blade servers and switches, x86-based Flexintegrated infrastructure systems, NeXtScale and iDataPlex servers and associatedsoftware, blade networking and maintenance operations
– Development, sales and marketing, finance, legal, integrated supply chain, operations, ITand manufacturing
– Service and support (maintenance)
IBM will retain its enterprise systems portfolio, including System z mainframes, PowerSystems, Storage Systems, Power-based Flex servers, and PureApplication and PureDataappliances
Lenovo and IBM plans to enter into a strategic collaboration
– Lenovo becomes IBM’s supplier of x86 server technology– Lenovo will license, OEM and resell IBM Storwize and tape storage technologies, General Parallel
File System, SmartCloud Entry, elements of the x86 system software portfolio, and the PlatformComputing portfolio
Until the transaction is completed, the companies will continue to operate independently
The transaction is expected to close later this year, subject to the satisfaction of regulatory requirements, customary closing conditions and anyother needed approvals. Subsequent local closings will occur subject to similar conditions, the information and consultation process, and localagreements in applicable countries. It is expected that once completed Lenovo will assume IBM’s x86 global server business including sales,development, marketing, service and support (maintenance) and related services, finance and certain parts of Integrated Supply Chain (ISC)operations
7© 2014 IBM Corporation
Who is Lenovo
Lenovo is a $34B company1
Seven different nationalities among its top ten executives The company is publicly traded on the Hong Kong Stock Exchange 60% publicly held, 32% held by Legend Holdings and 7% by it’s CEO
Lenovo is a truly global company 46,000 employees across 60 different countries WW Dual global head quarters Raleigh, NC USA – American executive leadership of it’s US operations Beijing, China
Major research centers in the United States, Japan, and China Manufacturing capabilities in the United States, China, India, and Mexico
Lenovo has experience in incorporating a former IBM business unit – PC Company Nine years post division sale to Lenovo they retain heavy investment WW They have taken #1 share in the PC segment WW
1FY 2013 Revenue Statement
8© 2014 IBM Corporation
More info:
Dave Weber: [email protected] / [email protected]
System x servers
Lenovo transition
9
Intel’s Commitment to Big Data Performance
Translating technology benefits into business impact.
The Haswell Processor: Performance with Energy Efficiency, up to 18 cores, many architectural enhancements
• DDR4: Lower power consumption and access latencies, higher frequency, higher B/W
• Use your processor’s capabilities to the fullest..You paid for them !
11
Working with Oracle and vendors to optimize Javafor Big Data. JD8u20, vectorization throughSuperWord
In Memory Analytics: Working with vendors tooptimize memory transformations and layout.Hadoop: 110sec/iterationSpark: 80 secs and then 1 sec per iteration
HDFS optimizations and benefits through Lustre’stransparent replication
Technology ingredients such as Intel® Data Plane Development Kit,Virtualization Enhancements and VMware* ESXi optimizations enable theNetwork Function Virtualization (NFV) vision
Storage & Network Trends• PCIe based storage such as Intel’s P3700 SSDs provide > 6X
performance gains• Just released Intel’s 40GBe NICs are ready for high bandwidth
applications such as network virtualization, trade view servers etc.• AMPS: Integrating In Memory streaming analytics with demanding
storage and networking bandwidth requirements:2X greater performance on HSW-EP versus IVB-EP !!
Intel’s contributions to emerging trends in Big DataSupplementing big cores with many cores.
- A trend to watch
STAC A2 Greeks Benchmark: Intel’s HSW–EP was 30% faster than IVB-EP
STAC A2 Greeks Benchmark: HSW-EP plusone Xeon PHI Coprocessor card was 22%faster than a system with two CPUs and 2GPUs. Demonstrated 46 % higher assetcapacity and 53% increase in higher pathscapacity
Follow on Xeon Phi to be more Big Dataenabled with higher memory capacity.
References:
12
• Today’s Speaker: Rama Karedla, Performance Architect, FSI [email protected]
• https://stacresearch.com//news/2014/09/08/stac-reports-intels-new-haswell-server-chip-and-without-xeon-phi-stac-a2
• BIGS004 - Accelerating Hadoop* Performance on Intel® Architecture BasedPlatforms
www.intel.com/idfsessionsSF
• BIGS001 - In-Memory Low Latency Analytics: Opportunities and ArchitectureTrends
www.intel.com/idfsessionsSF
DATS012 - Intel® Data Plane Development Kit: Open Source Foundations andVMware* Usages
www.intel.com/idfsessionsSF
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.
HPC Strategy for FSIEd TurkelGroup Manager, HPC Business DevelopmentHyperscale Business Group, HP Servers
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.14
Are you ready?
Accelerating pace of change requires a newstyle of IT
Advancing technologies
Cloud
MobilitySecurity
Big Data
Escalating demands
By 202030 billiondevices
40 trillion GBdata 10 million
mobile apps
8 billionpeople
…for
Changingconsumptionmodels
Evolvingecosystem
of ITproviders
IT requires moreperformance,
efficiency,sustainability
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.15
Global support and services | Best-in-class partnerships | Converged solutions
Convergence to accelerate IT servicedelivery
For virtualized and cloud workloads
HP BladeSystem HP OneView
Availability to function in real-time
For mission-critical environments
HP ProLiantscale-up
HP IntegrityNonStop
“DragonHawk” HP Integrity blades& Superdome
Intelligence to increase productivity
For core business applications
HP ProLiant ML HP ProLiant DLHP MicroServer
Density and efficiency to scale rapidly
For Big Data, HPC, and web scalability
Workload-optimized portfolio for better business outcomes
Commonmodularcompute
architectureHP ProLiant SL HP Moonshot HP Apollo
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.16
Delivering a complete HPC solution
Cloud
ServersCompute
StorageBig Data
AcceleratorsCompute, Viz
Network
Services
Power &Cooling
Management
ClientSystems
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.17
HAVEn – Big Data platform
HAVEn
Social media IT/OT ImagesAudioVideoTransactional
dataMobile Search engineEmail Texts
Catalog massivevolumes ofdistributed data
Hadoop/HDFS
Process andindex allinformation
AutonomyIDOL
Analyze atextreme scalein real-time
Vertica
Collect & unifymachine data
EnterpriseSecurity
PoweringHP Software+ your apps
nApps
Documents
hp.com/haven
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.18
Reinventing HPC todayto accelerate the world of tomorrow
HP Apollo familyOptimizing rack-scale computing for HPC
Acceleratingperformanceto speed up answers
Maximizingefficiencyfor sustainability and savings
UnleashingHPCto enterprises of any size
4x teraflops
per square foot
4x density per
rack per dollar
Years to daysfor new innovations
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.
Thank you
20 Copyright © 2012, Oracle and/or its affiliates. All rights reserved.
Oracle’s Big Data Platform
OracleExalytics
InfiniBand
OracleReal-TimeDecisions
OracleBig DataAppliance
OracleExadata
InfiniBand
OEP
DataWarehouse
HadoopOracle Big Data
Connectors
Oracle R dist
Oracle NoSQLDatabase
Oracle DataIntegrator
OracleAdvanced
Analytics(ORE/ODM)
OracleDatabase
Stream AcquireOrganize /Discover
AnalyzeVisualize /
Decide
Oracles DataIntegrator
Endeca/OBIEE
Flume
DataReservoir
ExternalData
Sources
Internal DataSources
Copyright © 2014, Oracle and/or its affiliates. All rights reserved. |
Oracle’s NoSQL on YCSB (Yahoo Cloud Serving Benchmark)
• SPARC T5-8 Flexible implementation deployment strategies– Use larger server with virtualization example above using Solaris Zones (1 zone/chip)– Use multiple T5-2 servers– Use Multiple-large servers
Oracle Confidential – Highly Restricted 21
SPARC also very efficient at Big Data: Oracle’s NoSQL
YCSB chip core Processor Ops/sOps/s
per chipOracle per chip
Advantage
Oracle T5-8 8 zones 8 128 3.6 SPARC T5 1,198,918 149.9k 3.5x
Cisco C240 M312node
24 192 2.9 E5-2690 1,028,868 42.9K 1.0
Oracle T5-8 4 zones 4 64 3.6 SPARC T5 636,765 159.2K
Cisco C240 M3 3node 6 48 2.9 E5-2690 302,153 50.4k
Copyright © 2014, Oracle and/or its affiliates. All rights reserved. |
Oracle’s Graph (Big Data Analytics)SPARC also very efficient at Big Data: Graph Algorithm
• SPARC 1.5x faster than E7 v2 per chip, great scalability– Graph is dependent delivered memory bandwidth
– BTE/s = Billion Traversed Edges per Second
Oracle Confidential – Highly Restricted 22
Graph chip Processor Problem Size2^30 & 2^31
PerfBTE/s
SPARC per chipAdvantage
Oracle T5-8 8 3.6 SPARC T5 31 1.67
Oracle T5-8 8 3.6 SPARC T5 30 1.73 1.6x faster than x86 E7 v2
X4-8 8 2.8 E7 v2 31 0.67
X4-4 4 2.8 E7 v2 30 0.54 Baseline
Copyright © 2014, Oracle and/or its affiliates. All rights reserved. |
SPARC: Leads in Uniform System Bandwidth
• SPARC fully-connected SMP, uniform memory bandwidth
• NUMA: local memory fast, but much slower bandwidth to other memory
– Non-local memory is slow chip-to-chip, often with multiple hops
23
IBM Power8/7+
Inter-chip bandwidth lines to scale
SPARC T5-8 x86 E7
Copyright © 2014, Oracle and/or its affiliates. All rights reserved. |
SPARC M7 Processor
• 32 SPARC Cores– 4th Generation CMT Core (S4)
– Dynamic, 8 Threads Per Core
• New Cache Organizations– Shared Level 2 Caches
– 64MB Shared Level 3 Cache
• DDR4 DRAM– +2TB Memory per Processor
– 2x-3x Memory Bandwidth
• 1 to 32 Processors SMP
• Technology: 20nm, 13ML
2.5x to 3.5x Performance of SPARC M6
4 cores 4 cores 4 cores 4 cores
4 cores 4 cores 4 cores 4 cores
SPARC M7
Fully-Connected 8 Processor SMP> 1 TB/s Delivered Bisection Bandwidth
[email protected]/solutions/big-data/
My NETWORKTrading
SYSTEMIs BETTER
THANYOURS
Trends driving innovation
● Switches silicon latencies are minimal and normalized
● Intelligence @ those expected latencies is now key
● Active feedback and intelligent control can reduce
trading platform latency, increase predictability and
profit
● Multi-tenancy economics are becoming more important
● Removal of dedicated security devices is a viable option
now
Multi tenancy: Why?
● Efficient use of costly infrastructure and real estate
● No-compromise virtualization of resources
● Removal of application or business silos
● Standardized operations and automation
● More performance, more redundancy, less
downtime due to human error
Multi tenancy: What technologies?
• VRF, NAT – virtualize the network, provide
isolation and security
• Programatic L2-4 control via API’s
• Application insight via in band and OOB system
monitoring
• Arista EOS: DANZ, LANZ, Burst Monitor, eAPI
Burst Monitor
Millisecond level reporting on interface burst activity with configurable intervals,
thresholds, and real time alerts. Provides application behavior insight and early
detection of burst based loss.
● Leverages high resolution ASIC counters
to provide real time burst detection
● Available on every port, RX and TX
● Provides burst insight without expensive
overlay infrastructure
● EOS provides programmable actions on
alerts
EOS BurstMonitor
Local logsSyslog
Trigger localor remote
action
Burst Monitor - Prototype in actionapl-7150S-24-1@12:09:35(config)#burstmonitor --help
Usage: burstmonitor [options] <interface number>
Options:
-h, --help show this help message and exit
-d, --debug print debug info
-r RXTOLERANCE, --rx-tolerance=RXTOLERANCE
RX burst size (% of bandwidth) which triggers syslog
messages (default=80)
-t TXTOLERANCE, --tx-tolerance=TXTOLERANCE
TX burst size (% of bandwidth) which triggers syslog
messages (default=80)
apl-7150S-24-1@12:06:24#show log | tail
Aug 15 12:06:28 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (254688B/1892us) detected on port 21
Aug 15 12:06:28 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (280460B/2051us) detected on port 21
Aug 15 12:06:28 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (250140B/1851us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (245592B/1796us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (259236B/1911us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (260752B/1939us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (541212B/2332us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (263784B/1946us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (272880B/1993us) detected on port 21
Aug 15 12:06:29 apl-7150S-24-1 ibm-21: %INTERFACE-4-RX_BURST_DETECTED: RX burst (269848B/1963us) detected on port 21
Security innovations - Why
• Many organizations still use firewall classdevices for some transaction flows
• NAT and deep inspection drive “intelligence” atthe exchange edge
• Replacement of these devices reduces latencyand increases capacity, but hard to meet“compliance” with security teams
Security innovations - What
Intelligent Bypass and NAT - bypass firewalls orapplication aware appliances to reduce latency,appliance cost, and pps constraints. Applicationhealth checking and dynamic NAT configuration.
Application Inspection - Dig deeper into theapplication layer to make intelligent decisionsbeyond the network header
Firewalls: Intelligent Bypass
● Direct flows based on applications, users, devices, content, threats, and
more
● Reduce end to end latency and accelerate trusted flows without
compromising security
● Protect firewall and server resources from oversubscription or attack
● Leverage firewalls in low latency environments to scrutinize suspicious traffic
● Size firewall resources for baseline traffic levels, letting the switch handle the
bulk or burst flows during traffic peaks
● Leverage L2-4 filtering and dynamic NAT with state sync* to bypass
firewalls completely where appropriate
*coming soon
Application Inspection● Inspect beyond the network header for decision
making and security
● Steer traffic based on a flexible fixed offset parser
● Make intelligent decisions on the application or data
type
○ Route financial data on content
○ Forward messaging layer/middleware on topic
tags
○ Selectively permit/filter/redirect any interesting
data up to 112 bytes
● Dynamically configure DPI intelligence via eAPI
● All done @ wire rate with standard forwarding
latency
© 2014 IBM Corporation