28
1 Presentation name Hybrid Technologies Extending Moore’s Law for Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk Tape SSNR Compressi on NF S SM B POSI X Swift /S3 HDF S Encryption SSD Disk Fast Disk Slow Disk AVAILABLE @FNAL THROUGH MAY 30

Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

1Presentation name

Hybrid Technologies Extending Moore’s Law forMachine Learning

Spectrum Scale

SSDDisk

FastDisk

SlowDisk

TapeSSNR

Compression

NFS

SMB

POSIX

Swift/S3

HDFS

Encryption

SSDDisk

FastDisk

SlowDisk

AVAILABLE @FNALTHROUGH MAY 30

Page 2: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

FEBRUARY 16, 2017

Page 3: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

NCSA:22,400 Servers716,800 cores20 hour run time

1 Billion Cell Reservoir

Solved 1000s of timesFaster…….

Page 4: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

APRIL 25th, 2017IBM-Nvidia Servers Achieve High-Performance Computing Milestone In Oil Industry

Pumpjacks operating at the Kern River Oil Field, in Bakersfield, California (Photo credit: AP Photo/Jae C. Hong, File)

In February, Exxon Mobil bragged it had reached a major computing breakthrough in its industry. Across 716,800 processors (that’s 32 processors on 22,400 servers) at the National Center for Supercomputing Applications’ Blue Waters supercomputer, the oil giant ran a billion cell model of a reservoir. Each one of those billion cells represents different characteristics of the reservoir — soil, fluid pressure, etc. The more cells there are, the more accurate the simulations for the reservoir are (and the more efficiently companies can extract oil). The results gave the company a data output that was thousands of times faster than normal reservoir simulations, Exxon boasted.

30 MINSKY SERVERS, 60U

60 x Power8 Processors = 480cores120 x P100s = 430,080cores

92 mins RUN TIME

<$2m

~1/1000th the number of Servers~1/2 the number of cores-1/10th the run time.

Page 5: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

5Presentation name

BUILT ON IBM POWER AND IBM ELASTIC STORAGE7,600 POWER NODES: 200Pflops• 2x IBM POWER CPUs

• 4-6x NVIDIA Tesla GPUs– NVLink connects GPUs at 80 GB/s

• 2x Mellanox EDR InfiniBand

• 800 GB NVMe storage

Spectrum Scale

SSD FastDisk

SlowDisk

TapeSSNR

Compression

NFS SMBPOSIX Swift/S3HDFS

Encryption

LLNL: 130PBytes of Storage: 1.2TB/s

130 Racks of Spinning Disk

Page 6: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

Unified Memory Access

<$70,000 with POWERAI Optimization

Page 7: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

4XThreads per core

4X Mem. Bandwidth

4XMore cache

These design decisions result in best performance for data centric workloads like: AI, ML, DL, Database, NoSQL, Big Data Analytics, OLTP

POWER8SMT8

x86Hyperthread

Parallel ProcessingPOWER8

pipe

Data flow

x86 pipe POWER8

x86 POWER8 + OpenPOWER

x86

Page 8: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

2.2ML:seeing2-20x Speed-uponIBM/NVLINK/P100:

2.2x

20x quote: Minds.ai benchmarks

Page 9: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2016 IBM Corporation

• All results are based on running Kinetica “Filter by geographic area” queries on data set of 280 million simulated Tweets with 1 up to 80 simultaneous query streams each with 0 think time. • Power System S822LC for HPC; 20 cores (2 x 10c chips) / 160 threads, POWER8 with NVLink; 2.86 GHz, 1024 GB memory, 2x 6Gb SSDs, 2-port 10 GbEth, 4x Tesla P100 GPU; Ubuntu 16.04. • Competitive stack: 2x Xeon E5-2640 v4; 20 cores (2 x 10c chips) / 40 threads; Intel Xeon E5-2640 v4; 2.4 GHz; 512GB memory 2x 6Gb SSDs, 2-port 10 GbEth, 4xTesla P100 GPU, Ubuntu 16.04. • Pricing is based on list pricing for S822LC for High Performance Computing http://www-03.ibm.com/systems/power/hardware/linux-lc.html on ibm.com and Dell C4130 priced at www.synnexcorp.com/us/govsolv/wp.../GSA70-GS-35F-0143R-Price-File-Dell.xlsx as of January 2017.

Unprecedented Database Performance with Tesla P100 GPUs –2.7X Better Performance at 61% Lower Cost per Query/Sec

• Accelerate the performance of Kinetica with 2.7X better performance than x86 accelerated solutions

– Power Systems S822LC with 4 Tesla P100s: 188,852 queries per hour

– Xeon E5-2640 v4 system with 4 Tesla P100s: 68,785 queries per hour

• POWER8 delivers this performance at 61% lower cost per transaction than x86

– Power Systems S822LC with 4x Tesla P100s: $421 per k query/hour

– Xeon E5-2640 v4 system with 4 Tesla P100s : $1081 per k query/hour

9

188852

68785

020000400006000080000

100000120000140000160000180000200000

POWER8 x86

Thro

ughp

ut (q

uerie

s/ho

ur)

IBM Power S822LC (20c/4x Tesla P100)

2x Xeon E5-2640v4 (20c/4x Tesla P100)

421

1081

0

200

400

600

800

1000

1200

POWER8 x86

$ pe

r 100

0 qp

h

2.7XMore

Throughput

61%LowerCost

IBM Power S822LC (20c/4x Tesla P100)

2x Xeon E5-2640v4 (20c/4x Tesla P100)

Page 10: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

FEEDING THE BEAST P100 = 5.3TFLops DB

Page 11: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

HealthcareDiagnosis/Treatment

FinancialFraudanalysis

PatternRecognition,Discovery

Retail Vision-enabledsolutions

Vision,Image,ObjectRecognitionNLP,Speech

On-Prem &CloudScale-outservers,CloudDeployment

DLFrameworksCAFFE,Torch,TensorFlow,Theano,Chainer,CNTK,MxNet

IndustrySolutions

MLDLBuildingBlocks,Services,APIs

SoftwareLibraries,Frameworks,Platforms

Infrastructure

...Knowledge

Representation

DataLayer AcceleratedDBsNoSQLDBs Streams

Recommendation,Prediction

HDFS

AcceleratorsGPUs,FPGAs,CAPI Flash,DLAccelerators

UniquePowerTechnologiesNVLINK,CAPI,OpenCAPI

MLLibrariesSparkML,SparkR,SciKit,Numpy,SystemML,Mahout,H2O

Cloud: IBM(DSX,DLaaS,MLaaS,Watson),Google,Nimbix,Microsoft

DistributedComputing:Hadoop,Spark,MPI

...

LegacyDBs

MathLibraries:OpenBLAS/MASS,ATLAS,ESSL,LAPACK,cuDNN,cuBLAS,cuSparse,…

DiscoveryHEPhysicsNuclearPhysicsSimulations

GPFSObject

OnPrem: IBMDSX/ML,PowerAI Pro

MachineLearningEcosystemforMinsky:Data,QueryTools,ML,InferenceEngine,RealTimeDecisions

HEPhysics

Page 12: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

12

PowerAI:AllEnabledandOptimizedforMINSKY…

Caffe NVCaffe TorchIBMCaffe

DL4JTensorFlow

OpenBLAS

Theano

Deep Learning Frameworks

Accelerated Servers and

Infrastructure for Scaling

SpectrumScale:High-SpeedParallel

FileSystem

ScaletoCloud

ClusterofNVLinkServers

Bazel DIGITSNCCLDistributedFrameworks

Supporting Libraries

ChainerFREE

Page 13: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

13Presentation name

ApplicationsOptimizedforPower/NVIDIAforCORALDomain Application Methods PI Institution Relatedto INCITE/ALCC RelatedtoSciDAC

Astrophysics FLASH Grid,AMR BronsonMesser ORNL Zingale SciDAC II

Chemistry DIRAC Particle, LA LucasVisscher VUA Dixon

ClimateScience ACME(N) Unstr Mesh DavidBader LLNL Taylor SciDAC III

Engineering RAPTOR Kokkos Joseph Oefelein SNL Oefelein SciDAC II

MaterialsScience QMCPACK MC PaulKent ORNL Kent,Ceperley SciDAC III

NuclearPhysics NUCCOR Particle Gaute Hagen ORNL Vary SciDAC III

PlasmaPhysics XGC(N) PIC,PETSc CSChang PPPL Chang SciDAC III

Seismic Science SPECFEM Unstr Mesh Jeroen Tromp Princeton Tromp

Astrophysics HACC(N,A) Grid SalmanHabib ANL Habib SciDAC III

Biophysics NAMD(N) Particle KlausSchulten UIUC Klein,Schulten,Tajkhorshid SciDAC II

Chemistry NWCHEM(N) Particle,LA KarolKowalski PNNL Dixon,Sumpter SciDAC III

Chemistry LSDALTON Particle,LA Poul Jørgensen Aarhus Jørgensen

PlasmaPhysics GTC(N) PIC Zhihong Lin UCI Lin SciDAC III

N:NERSCapplication;A:ALCFapplication

Page 14: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2016 IBM Corporation

YCSB running MongoDB on POWER8 delivers leadership performance and 1.68X better price-performance than Intel Xeon E5-2690 v4 Broadwell

IBM Power S822LC for Big

Data(20-core, 128GB)

HP DL380

(28-core, 128GB)

Server price-3-year warranty

$11,581 $16,626

System Cost-Server + RHEL OS + MongoDB Annual Subscription

$24,870($11,581 + $1,299 + $11,990)

$29,915($16,626 + $1,299 +

$11,990)

MongoDB YCSB(total operations per second)

288,824 ops 205,951 ops

Op per Sec / $ 11.6 ops/$ 6.9 ops/$1.68X

better

2XBetter Price-Performance

-IBM GUARANTEED-

30%Lower HW costs and

maintenance

40%More Performance

per Server

@

•Based on IBM internal testing of single system and OS image running Yahoo Cloud Services Benchmark (YCSB) 0.6.0, 1M record workload at 50/50 read/write factor. Results valid as of 8/24/16. Conducted under laboratory condition, individual result can vary based on workload size, use of storage subsystems & other conditions.• IBM Power System S822LC for Big Data; 20 cores (2 x 10c chips) / 160 threads, POWER8; 2.9 GHz, 128 GB memory, MongoDB 3.3.8 RHEL 7.2. Competitive stack: HP Proliant DL380, 28 cores (2 x 14c chips) / 56 threads; Intel E5-2690 v4; 2.6 GHz; 128 GB memory, MongoDB 3.3, RHEL 7.2 . Both servers running favor performance mode and priced with 2 x 1TB SATA 7.2K rpm HDD, 1 Gb 2-port, 1 x 16gbps FCA. Configurations represent the specific processor running the MongoDB server on 1 socket & the YCSB application workload on the 2nd socket. IBM Flash 900 storage was used on both server for testing.•Pricing is based on: S822LC for Big Data http://www-03.ibm.com/systems/power/hardware/linux-lc.html , and HP DL380 https://h22174.www2.hp.com/SimplifiedConfig/Index MongoDB https://www.mongodb.com/compare/mongodb-oracle Page: 6

Page 15: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

HOW IS STORAGE KEEPING UP WITH DATA?

Page 16: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© Copyright IBM Corporation 201716

NAS STORAGE: INTEGRATED GPFS - Erasure Coding, - 512KB-16MB Blocksize- Inode Data Caching- In Memory MetaData- Auto Tiering; Flash to Tape~$300/TB

Spectrum Scale

ESS

New! Model GL2S: 2 Enclosures, 14U

166 NL-SAS, 2 SSD

New! Model GL4S: 4 Enclosures, 24U

334 NL-SAS, 2 SSD

New! Model GL6S: 6 Enclosures, 34U

502 NL-SAS, 2 SSD

ESS 5U84 Storage

ESS 5U84 Storage

Max: .6PB usable Max: 1.1PB usable Max: 1.4PB usable Max: 2.6PB usable Max: 2.25PB usable Max: 4PB USABLE

Model GL2: 2 Enclosures, 12U

116 NL-SAS, 2 SSD

ESS 5U84 Storage

ESS 5U84 Storage

ESS 5U84 Storage

ESS 5U84 Storage

Model GL6:6 Enclosures, 28U

348 NL-SAS, 2 SSD

Model GL4: 4 Enclosures, 20U

232 NL-SAS, 2 SSD

ESS 5U84 Storage

ESS 5U84 Storage

ESS 5U84 Storage

ESS 5U84 Storage

ESS 5U84 Storage

ESS 5U84 Storage

36 GB/s

25 GB/s

17 GB/s

12 GB/s8 GB/s

24 GB/s

All performance is dependent on customer network and workload

See performance planning spreadsheet tool for your specific environments

Page 17: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© Copyright IBM Corporation 201717

Updated ESS Performance / Capacity spreadsheet

GL6S = 36 GB/s

GL6 = 25 GB/sGL4S = 24 GB/s

GL4 = 17 GB/s

GL2S = 12 GB/s

GL2 = 8 GB/s

Spectrum ScaleESS

IBM and Business Partner Confidential until Announcement

Page 18: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2016 IBM Corporation

Neo4j on POWER8 with CAPI enabled acceleration1.61X the price-performance versus Intel Xeon E5-2650 v4 with NVMe

IBM Power S822LC(20-core, 128GB)

HP DL380 Gen9(24-core, 128GB)

Server price*-3-year warranty

$19,123 $16,911

Mixed graph transaction Workload(total operations per second)

711 390

1.61XPrice-Performance

1.82XPerformance per Server

• Based on IBM internal testing of single system and OS image running mixed graph transaction s based on 200 GB data model internal IBM and Neo4j workload. Conducted under laboratory condition, individual result can vary based on workload size, use of storage subsystems & other conditions. Data as of October 19, 2016

• IBM Power System S822LC; 20 cores (2 x 10c chips) / 160 threads, POWER8; 128 GB memory (16 x 8GB), 1.6 TB CAPI NVMe adapter , Neo4j 3.0.4, Ubuntu 16.04. Competitive stack: HP Proliant DL380 Gen9; 24 cores (2 x 12c chips) / 48 threads; Intel E5-2650 v4; 128 GB memory,(16 x 8GB), 1.6 TB NVMe adapter, Neo4j 3.0.4, Ubuntu 15.10.

* Pricing is based bundled pricing for S822LC with Integrated CAPI Flash card (IBM ordering system) and HP Web price https://h22174.www2.hp.com/SimplifiedConfig/Index

18

Page 19: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

19Presentation name

CORAL PROJECT: Summit will replace Titan as the OLCF’s leadership supercomputer in 2019

• Dramatical fewer nodes

• Much more powerful nodes

• Much more memory per node and total system memory

• Faster interconnect

• Much higher bandwidth between CPUs and GPUs

• Much larger and faster file system

Feature Titan SummitApplication Performance Baseline 5-10x Titan

Number of Nodes 18,688 ~4,600

Node performance 1.4 TF > 40 TF

Memory per Node 38GB DDR3 + 6GB GDDR5 512 GB DDR4 + HBM

NV memory per Node 0 800 GB

Total System Memory 710 TB >6 PB DDR4 + HBM + Non-volatile

System Interconnect (node injection bandwidth) Gemini (6.4 GB/s) Dual Rail EDR-IB (23 GB/s)

Interconnect Topology 3D Torus Non-blocking Fat Tree

Processors 1 AMD Opteron™1 NVIDIA Kepler™

2 IBM POWER9™6 NVIDIA Volta™

File System 32 PB, 1 TB/s, Lustre® 250 PB, 2.5 TB/s, GPFS™

Peak power consumption 9 MW 13 MW

Page 20: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2016 IBM Corporation

YCSB running MongoDB on POWER8 delivers leadership performance and 1.68Xbetter price-performance than Intel Xeon E5-2690 v4 Broadwell

IBM Power S822LC for Big

Data(20-core, 128GB)

HP DL380

(28-core, 128GB)

Server price-3-year warranty

$11,581 $16,626

System Cost-Server + RHEL OS + MongoDB Annual Subscription

$24,870($11,581 + $1,299 + $11,990)

$29,915($16,626 + $1,299 +

$11,990)

MongoDB YCSB(total operations per second)

288,824 ops 205,951 ops

Op per Sec / $ 11.6 ops/$ 6.9 ops/$1.68X

better

2XBetter Price-Performance

30%Lower HW costs and

maintenance

40%More Performance

per Server

@

•Based on IBM internal testing of single system and OS image running Yahoo Cloud Services Benchmark (YCSB) 0.6.0, 1M record workload at 50/50 read/write factor. Results valid as of 8/24/16. Conducted under laboratory condition, individual result can vary based on workload size, use of storage subsystems & other conditions.• IBM Power System S822LC for Big Data; 20 cores (2 x 10c chips) / 160 threads, POWER8; 2.9 GHz, 128 GB memory, MongoDB 3.3.8 RHEL 7.2. Competitive stack: HP Proliant DL380, 28 cores (2 x 14c chips) / 56 threads; Intel E5-2690 v4; 2.6 GHz; 128 GB memory, MongoDB 3.3, RHEL 7.2 . Both servers running favor performance mode and priced with 2 x 1TB SATA 7.2K rpm HDD, 1 Gb 2-port, 1 x 16gbps FCA. Configurations represent the specific processor running the MongoDB server on 1 socket & the YCSB application workload on the 2nd socket. IBM Flash 900 storage was used on both server for testing.•Pricing is based on: S822LC for Big Data http://www-03.ibm.com/systems/power/hardware/linux-lc.html , and HP DL380 https://h22174.www2.hp.com/SimplifiedConfig/Index MongoDB https://www.mongodb.com/compare/mongodb-oracle Page: 6

Page 21: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2016 IBM Corporation

IBM TS4500 Tape Library: STK T10000 upgrade

Lee JesionowskiLead Architect – Tape Automation Products

IBM Confidential

Page 22: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

22 © 2016 IBM Corporation

Branded Tape Automation Revenue Share (Rolling 4 Quarters Through 2Q16)

Branded Tape Enterprise AutomationRevenue Share

Source: IDC Branded Tape Pivot Q2CY16

Dell12.5%

HP18.3%

IBM28.6%

Oracle15.8%

Others2.2%

Overland2.3%

Quantum14.9%

Spectra Logic5.4%

IBM74.5%

Oracle25.5%

HP6.4%

IBM46.7%

Oracle25.5%

Others0.5%

Quantum9.2%

SpectraLogic11.7%

Branded Tape Midrange (i.e. LTO) Automation 500+ slots Revenue Share

IBM Makes Tape:- 100% of LTO7- 100% of IBM Enterprise

Page 23: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

23 © 2015 IBM Corporation

TS11XX Investment Protection - Media Re-use and Drive Model Upgrades

3592 J1ADrive

TS1120Drive

TS1130Drive

TS1140Drive

JA Cartridge Read Backwards all JA formats (N-3)

Cartridge

300 GB

500 GB

700 GB

JB Cartridge

Read Backwards all JB formats (N-2)700 GB

1000 GB Reads/Writes JB 1 TB format (N-1)

1600 GB Up-formats/Reads/Writes JB to 1.6 TB

4000 GBJC Cartridge

Reads and Writes JC media at 4 TB @ 250MB/s

7 TB @300MB/s

JC re-usable on TS1150/60 drives

Summary • Drive model upgrades allow the upgrade of drive to the next generation at modest expense • Up-formatting media provides the ability to use legacy media at higher capacity and performance in the new drive, preserving and enhancing the media investment

TS1150Drive

Drive Model Upgrades

Media Re-use / Up format

Page 24: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2016 IBM Corporation

Thank you!

Page 25: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© Copyright IBM Corporation 2017© Copyright IBM Corporation 2017

Spectrum Scale Capacity Licensing includes multiple clusters*

Cross-cluster mount:http://www.ibm.com/support/knowledgecenter/STXKQY_4.2.0/com.ibm.spectrum.scale.v4r2.adv.doc/bl1adv_admmcch.htm

bb

Spectrum Scale NSD data servers all w/ Capacity License

Spectrum Scale servers asmount point for remote clusters

Spectrum Scale quorum nodes

Spectrum Scale Protocol Nodes

Other Spectrum Scale NSD Servers

Spectrum Scale NSD data servers all w/ Capacity License

High speed data LAN

Spectrum Scale data / storage cluster #1

CIFS, NFS users

User LAN

(Ethernet)

Spectrum ScaleClient licenses, quorum nodes in Compute cluster also covered* by

Capacity license, when doing remote mount to data / storage cluster

User LAN(Ethernet)

Spectrum Scale quorum nodes

Spectrum Scale client / compute / application cluster 100s / 1000’s of compute nodes….

Object users

Hadoop users

Transparent Cloud Tiering users

Spectrum Scalecompute clients

Spectrum Scalecompute clients

Transparent Cloud TieringProtocol Nodes

25 *Terms and conditions apply

Goal is to sell new clusters with Spectrum Scale Data

Management Edition capacity licenses

Page 26: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

© 2017 IBM Corporation

MongoDB running on POWER8 Price-Performance Guarantee

• •IBMPowerSystemsguaranteesthePowerS822LCforBigDatasystembuiltwithPOWER8deliversatleasta2Xprice-performanceadvantagevs.x86basedserverswhenrunningacustomerapplication/workloadbasedonMongoDB.

2Xprice-performancemeansthatthecustomer'sdocumentedthroughputperformanceontheS822LCPOWER8dividedbythepriceofthesystemwillbeatleast2timeshigherthanthecustomer'sdocumentedthroughputperformanceonthex86basedsystemdividedbytheprice of thecomparable

x86system.

EX: If transactions per second on the S822LC are 20,000 and 10,000 on the x86 based system, while the price of the S822LC is $10,000, and the price ofthe x86 based system is $10,000, then the Throughput Performance Per Price would be exactly 2 times higher and the guaranty would be met."

Notes:1. Client’s POWER8 Machine and the x86 Machine must be running at similar utilization rates.2. Client’s POWER8 Machine’s system performance cannot be constrained by I/O subsystem. Specifically, the I/O subsystem on the POWER8 Machines must achieve greater than or equal I/O bandwidth and operations per second than the x86 Machine.3. Client’s POWER8 Machine’s physical memory must be the same or greater than the physical memory on the x86 Machine4. Client is responsible for demonstrating comparable real-world representative workload between the POWER8 Machine and the x86 Machine through the use of the IBM provided tools and comparable tools on x86 systems.5. 1.8x guarantee is based on a list price for x86 (Dell, Cisco, HP or Lenovo) and the IBM S822LC for Big Data.

The IBM Power S822LC for Big Data server (20-core/2.92 GHz 128GB memory, 4 TB SATA Storage) must be purchased from IBM or an authorized IBM Business Partner prior to June 30, 2017. The guarantee period is valid for three (3) months from the date of purchase. The x86 based systems must be comparably configured branded servers from Cisco, Dell, HP, or Lenovo and the client is responsible for all MongoDB licenses.

2X throughput performance per price means that the customer's documented throughput performance on the S822LC POWER8 system based on either queries, operations or transactions per second divided by the price of the such system will be at least 2 times higher than the customer's same documented throughput performance on the x86 based system divided by the price of such comparable x86 system.

Remediation: IBM will provide additional performance optimization and tuning services consistent with IBM Best Practices, at no charge. If unable to reach guaranteed level of price-performance, IBM will provide additional equally configured systems to those already purchased to reach the guaranteed level of price-performance.

Page 27: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

EnterpriseDB on Power Systems Price-Performance Guarantee

27 © Copyright EnterpriseDB Corporation, 2017. All Rights Reserved.

IBMPowerSystemsguaranteestheS822LCforBigDatasystembuiltwithPOWER8deliversatleasta1.8Xprice-performanceadvantageversusx86basedserverswhenrunningavirtualizedcustomerapplication/workloadbasedon

EnterpriseDB Postgres9.5.

1.8Xprice-performancemeansthatthecustomer'sdocumentedthroughputperformanceontheS822LCPOWER8dividedbythesumofthepriceofthesystemandassociatedEnterpriseDB licenseswillbeatleast1.8timesthatofthe

customer'sdocumentedthroughputperformanceonthex86basedsystemdividedbythesumofthepriceofthecomparablex86systemandassociatedEnterpriseDB licenses

EX:IftransactionspersecondontheS822LCare18,000and10,000onthex86basedsystem,whilethepriceoftheS822LCandassociatedEnterpriseDB licensesis$10,000,andthepriceofthex86basedsystemandassociated

EnterpriseDB licensesis$10,000,thentheThroughputPerformancePerPricewouldbeexactly1.8timesadvantagedandtheguarantywouldbemet."

Notes:1. Client’s POWER8 Machine and the x86 Machine must be running at similar utilization rates. Eligible Machine and the Compared Machine must be partitioned with at least 4 equal sized partitions.2. Client’s POWER8 Machine’s system performance cannot be constrained by I/O subsystem. Specifically, the I/O subsystem on the POWER8 Machines must achieve greater than or equal I/O bandwidth and operations per second than the x86 Machine.3. Client’s POWER8 Machine’s physical memory must be the same or greater than the physical memory on the x86 Machine4. Client is responsible for demonstrating comparable real-world representative workload between the POWER8 Machine and the x86 Machine through the use of the IBM provided tools and comparable tools on x86 systems.5. 1.8x guarantee is based on list price (for the x86 based server (Dell, Cisco,or HP and list price for the IBM S822LC for Big Data.6. EnterpriseDB Postgres Advanced Server 9.5 license are priced at $1750 per core - EDB 9.5 http://www.enterprisedb.com/products-services-training/subscriptions-power

The IBM Power S822LC for Big Data server (20-core/2.92 GHz 256GB memory, 4 TB SATA Storage) must be purchased from IBM or an authorized IBM Business Partner prior to March 31, 2017. The guarantee period is valid for three (3) months from the date of purchase. The x86 based systems must be comparably configured branded servers from Cisco, Dell, HP, or Lenovo and the client is responsible for all EnterpriseDB licenses.

1.8 X price-performance means that the customer's documented throughput performance on the S822LC POWER8 divided by the sum of the price of the system and associated EnterpriseDB licenses will be at least 1.8 times that of the customer's documented throughput performance on the x86 based system divided by the sum of the price of the comparable x86 system and associated EnterpriseDB licenses

Remediation: IBM will provide additional performance optimization and tuning services consistent with IBM Best Practices, at no charge. If unable to reach guaranteed level of price-performance, IBM will provide additional equally configured systems to those already purchased to reach the guaranteed level of price-performance.

Page 28: Hybrid Technologies Extending Moore’s Lawfor Machine Learning · Hybrid Technologies Extending Moore’s Lawfor Machine Learning Spectrum Scale SSD Disk Fast Disk Slow Disk SSNR

28Presentation name

Spider 3 @ OLCFSpider 3 is a center-wide single namespace POSIX file system to serve all OLCF resources, eliminating data islands and enabling seamless data sharing between resources

• Built on IBM’s Elastic Storage Server and uses Spectrum Scale (formerly known as GPFS) parallel filesystem technology utilizing GPFS Native RAID with 8+2 redundancy

• Provides a usable capacity of 250 PB• Performs at an aggregate sequential peak read/write bandwidth of 2.5 TB/s• Performs at an aggregate random peak read/write bandwidth of 2.2 TB/s• Provides rich metadata performance; single directory parallel create rate of 50,000/s• Provides rich interactive performance; @32 KiB I/O 2.6 million IOPs• Disk-based, with tens of thousands of disks• Connected to OLCF’s SION 3 SAN with IB EDR • Will also serve as the Summit Burst Buffer sink and source on the end-to-end I/O path