Big Data Everywhere Chicago: High Performance Computing - Contributions Towards Big Data (HPC)

Sharan Kalwani
/ sharan.kalwani@acm.org
www.linkedin.com/sharankalwani
1

Outline
o History of Supercomputing *
o Technologies
o Modern Day HPC:
o Current State of the Art
o Peering beyond the Horizon:
o Next Set of technologies
* aka High Performance Computing (HPC)

History
 Computing Demand:
 driven by needs far beyond contemporary capability
 Early adopters: (1970s)
LANL (Los Alamos National Lab) and
NCAR (National Center for Atmospheric Research)
 Characteristics: domain specific needs
 Features: High Speed Calculations: PDE, Matrices
 1972: Seymour Cray (CDC, Cray Research Inc.)
 1st Model Cray-1

History
 Cray-1 Characteristics: (1975-1976)
 64 bit word length
 12.5 nanosecond clock speed
 80 MHz
 “original” RISC
 1 clock == 1 instruction
 Vector instruction set, true multiplier effect, single instructions, multiple data
 Matrix operations, pipelining,
 included add+multiply!
 memory <> processor balance
 Cray-1, Cray-XMP, Cray-YMP, Cray-2, Cray 3

History
 Enter the domain of MPP
 Massively parallel processors
 Introduction of Torus architectures
 Seen these days in some offerings
 Cray T3 D/E/F….(1st machine to break
1,000,000,000 calculations/sec barrier)

Cray T3 architecture (logical) circa 1993, looks a lot like a cluster, eh?

Hardware Contributions: _Phase_ 1
 Profusion of technologies:
 RISC inspiration (1 clock cycle → 1 instruction)
 Solid State Disk – recognized the need for keeping CPU busy all the time
 multi-core software – coherence + synchronization
 De-coupling of I/O from compute
 Massive memory
 I/O technologies – HiPPI (high speed parallel interface)
 Visualization driver
 Chip set design, ECL -> CMOS integration
 Parallel processing software foundation -> MPI

Solid State Disk (grandpa USB stick)
 The first CRAY X-MP system had SSD in 1982.
 Designed for nearly immediate reading and writing of very large data files.
 Data transfer rates of up to 1.250 GBytes / second,
 Far exceeding *any* other data transfer devices in its time.
 SSDs offered in sizes of 64, 128, or 256 million bytes of storage.
 The hole in the cabinet was to attach a very high speed (VHISP) data channel to an SSD.
 Link referred to as the "skyway."
 Via software, the SSD is logically accessed as a disk unit.
SSD driver ~ 1200 lines of C code!

History Marches on….
 Battle of technologies:
 Silicon v. Gallium Arsenide
 Vector v. Killer Micros
 Accelerated Strategic Computing Initiative (mid 90s) ASCI project changed directions for everyone

Speed
 Everybody was focused on clock speed
 Or Floating Point Operations (FLOPS/sec)
 @12.5 ns clock speed  80 million Flops/sec (peak)
 Leading to the famous Macho FLOPS race:
 USA v Japan (90s)
Megaflops  Gigaflops (1000 MF)
Gigaflops  Teraflops (1000 GF)
Teraflops  Petaflops (1000 TF)
In 2018 the industry expects an ExaFlop machine!

Speed
 First GigaFlop/sec* System
 Cray YMP and Cray-2
 First TeraFlop/sec System
 Sandia National Lab ASCI “Red” (Intel)
 First PetaFlop/sec System
LANL “RoadRunner” (IBM)
 First ExaFlop/sec System
???
* SUSTAINED!

Is anyone keeping score?
 Birth of the Top500 list
 1993 – Dongarra, Strohmaier , Meuer & Simon
 Linear Algebra Package (LINPAK) basis
 Offshoots:
 Green500 (power efficiency)
 Graph500 (search oriented – little to no floating point computation)
 @SC13 a new replacement metric has been proposed

Is anyone keeping score?
We will return to this …..

Cluster Growth propelled by HPC

Track Record of Linux versions in HPC
•See also Linux Foundation report
•http://www.linuxfoundation.org/publications/linuxfoundation/top500_report_2013

HPC top500 - factoids
•Current #1 system has 3,120,000 cores
–Located in China, called Tianhe-2 “Milky Way”
–Peak speed of 33.9 PetaFLops/second (quadrillions of calculations per second)
–Needs 17.8 MW of power
•Current #2 system @ ORNL (US Government DoE) in Tennessee
–Has 560,640 cores, called Titan
–Peak speed of 17.6 PetaFlops/second
–Needs 8.21 MW of power

•Tianhe-2

•Titan (http://www.ornl.gov/info/press_releases/get_press_release.cfm?ReleaseNumber=mr20121029-00)
•

•Titan

Operating Systems: History
 Early HPC OS:
 tiny assembled loaders
 CTSS (Cray Time Sharing Systems) - LTSS
 CRAY Operating Systems (COS)
 CRAY UNIX ((UNICOS)
 mk/Kernel – CHORUS
 Beowulf cluster – Linux appears

Linux Contributions: History
 Linux – attack of the killer micros, 1992

 NOW – Network of workstations, 1993

Linux Contributions: History 1993-1994
 133 nodes – Stone Supercomputer
 First Beowulf cluster
 Concept pioneered at NASA/Caltech
 Thomas Sterling and Donald Becker

•Beowulf

•Beowulf
•NASA
•LSU
•Indiana University
THOMAS STERLING

•Beowulf Components:
–Parallel Virtual Machine (PVM) – U Tennesse
–Message Passing Interface (MPI) – several folks
–Jack Dongarra,Tony Hey and David Walker
–Support of NSF and ARPA
–Today we have the MPI Forum
–MPI 2 and now MPI 3
–OpenMPI, MPICH, etc
–Future Pthreads and OpenMP,

HPC and Linux
•Beowulf Attributes (or cluster features):
•Open Source
•Low Cost
•Elasticity
•Equal Spread of work (seeds of cloud computing here!!)
•These days the Linux kernel can handle 64 cores! HPC pushes this limit even further….

HPC and Linux pushing the boundaries
•File systems:
–Large number of high performance file systems
–Lustre, now in version 2.5
–Beats HDFS several times over!!
–You can host HDFS over many HPC filesystems for massive gains

Typical Stack
Pick Distro – Linux based (usually Enterprise class)
Hardware

Hardware Contributions: _Phase_ 2
 Profusion of technologies:
 In-memory processing, many HPC sites implemented this
 1992 built special systems for the use in cryptography using these technigques
 Graph traversal systems – now available as appliances by HPC vendors
 Massive memory : single memory systems over several TB in size
 Infiniband interconnects: hitting 100 Gbits/sec switches you can buy them now
 Parallel processing software foundation -> replacements for MPI stack being worked on

Modern Day HPC
•Building the ExaScale machine:
–Exascale is 1 quintillion calculations/second
–1000x Petaflops/sec
–Also Known as 10^18 (hence 2018 projections)
–1, 000, 000, 000, 000, 000, 000 floating point calculations/second (sustained)
–How to feed this monster?

Modern Day HPC
•Solutions for the ExaScale monster:
•Inevitably Big Data community should watch/support/benefit issues we are tackling now:
–Memory matters!
–Resiliency in software
–Robustness in hardware
–Co-Design critical
–Power Consumption and Cooling (estimate several megawatts w/ present day approaches)
–Utterly new architectures needed

Applications: What did we use all this for?
Weather
Automotive and Aerospace Design
Traditional Sciences
Energy (Nuclear, Oil & Gas)
Bioinformatics
Cryptography
 and……big data

Applications: traditional HPC
Automotive & (similar) Aerospace Design
o Car Crash Analysis – prime usage, 50%
o Each physical crash test costs $0.5 million
o Virtual Prototype test - $1000 (or less)

Applications: The real deal vs. HPC
•NHTSA requires physical validation
•Before total crash tests cost a total of $100 million/year
•Limited to a small suite: 12 tests
•Today we can do over 140+ different tests (for each vehicle) and with:
–Less cost (we instead increased the # of tests!)
–Faster response (5 years v 12 months)
–Many more design iterations (Hundreds v 10)

Technology March! Or why simulation matters?
•Increasing resolving power - greater fidelity problem
•Decreasing product design turnaround times
•Increase cost-effectiveness relative to experiment & observation
•Reducing uncertainty
•Ramping up the ease of use by non-experts
•Powerful tool in resolving scientific questions, engineering designs, and policy support
•Co-execution environments for simulation, large-scale data enabled science, and scientific visualization
•Simple: Better Answers thus delivering an…..
–Attractiveness to the creative and entrepreneurial classes
–Straightforward case for national economic competitiveness!!!

HPC is indispensable!
•Establish Capability
•Enable Adding of Complexity
•Gain a real and better Understanding
•And do not forget all that data!
•How do we tie it in?......

Approaching the eXtreme Scale
•Current paradigm: Simulation lots of equations which mimic or model actual situations “Third” Paradigm
•-------------------------------------------------------------------------------------------
•Operate without models (Big Data) “Fourth” Paradigm

Operate without models (Big Data)
•BEFORE….. * NOW/FUTURE….
Models/Theory
Models/Theory
DATA

Best Example….. tell us what we do not know!
•Recent Success:
•Solar observations (actual data)
•Unknown Surface Perturbations or Energy
•Could not be explained by all classical models
•Resorted to automated machine learning driven alternate search
•Answer: Solar Earthquakes and Thunderclaps, classic acoustic signature!
•New profession: Solar Seismologists!

Trend began a decade+ ago…. http://research.microsoft.com/en- us/collaboration/fourthparadigm/4th_paradigm_book_complete_lr.pdf

Everyone is seriously interested….. http://science.energy.gov/~/media/ascr/ascac/pdf/reports/exascale_subcommittee_report.pdf.

A Peek at the Future…….
•Yes…we should definitely care
•The bigger and more relevant questions are:
–What architecture? What programming model?
–Power consumption will dominate
•Currently 4 approaches:
–Stay the course ? Not!!
–All GPGPU based?
–ARM based?
–Quantum Computing ??

GPGPU perspective….
G1
G2
G3
G4
8-Cores
8-Cores
16-Core Server Node
Multi-GPU Acceleration of a 16-Core ANSYS Fluent Simulation of External Aero Xeon E5-2667 CPUs + Tesla K20X GPUs
2.9X Solver Speedup
CPU Configuration
CPU + GPU Configuration
Click to Launch Movie

–GPGPU
GDDR
GDDR
DDR
DDR
GPU
I/O Hub
PCI-Express
CPU
Cache
1
2
3

–GPGPU based?
–http://www.anl.gov/events/overview-nvidia-exascale- processor-architecture-co-design-philosophy-and- application-results
–Echelon
–DragonFly
–http://www.nvidia.com/content/PDF/sc_2010/theater/Dally_SC10.pdf

–ARM or Pi based?
–http://coen.boisestate.edu/ece/files/2013/05/Creating.a. Raspberry.Pi-Based.Beowulf.Cluster_v2.pdf

Quantum Computing…….
•D-WAVE systems installed at NASA Ames lab
•Uses a special chip, 512 bit “Vesuvius”
•Uses 12 KW of power
•Cooled to 0.02 Degrees K (100 times colder than outer space)
•RF shielding

•Shor’s Integer Factorization Algorithm
•Problem: Given a composite n-bit integer, find a nontrivial factor.
–Best-known deterministic algorithm on a classical computer has time complexity exp(O( n1/3log2/3 n)).
•A quantum computer can solve this problem in O( n3 ) operations.
Peter Shor
Algorithms for Quantum Computation: Discrete Logarithms and Factoring
Proc. 35thAnnual Symposium on Foundations of Computer Science, 1994, pp. 124-134

•Classical: number field sieve
–Time complexity: exp(O(n1/3 log2/3 n))
–Time for 512-bit number: 8400 MIPS years
–Time for 1024-bit number: 1.6 billion times longer
•Quantum: Shor’s algorithm
–Time complexity: O(n3)
–Time for 512-bit number: 3.5 hours
–Time for 1024-bit number: 31 hours
•(assuming a 1 GHz quantum machine)
See M. Oskin, F. Chong, I. Chuang
A Practical Architecture for Reliable Quantum Computers
IEEE Computer, 2002, pp. 79-87

What I will be looking into…….
•Julia
•Programming Environment which combines *all* the elements of:
–R (express data handling)
–Scientific and Engineering process (e.g. MATLAB like)
–Parallel processing and distributed computing functional approaches (similar to Scala, Erlang and others)
–Python and other integration packages already there
–Happy marriage of several arenas
– Now in early release
•Feel free to contact or follow up with me on this

SUMMARY: Core Competencies Across HPC
Core Competencies
Extreme scale
Architecture
Compute
I/O
Memory
Storage/data management
Tera, Peta, Exabytes….
Visualization and analytics
Fast fabrics
Future architectural direction
Parallelism to extreme parallelism
Multi core
Programming models
Big Data
Models, applications, applied analytics
Structured, unstructured data types

The need for a new discipline: HPC experts + Domain Expertise == Simulation.Specialists
Core Competencies
Where would this Computational Specialist work?
Extreme scale
Architecture
Compute
I/O
Memory
Storage/data management
Tera, Peta, Exabytes….
Visualization and analytics
Fast fabrics
Future architectural direction
Parallelism to extreme parallelism
Multi core
Programming models
Big Data
Models, applications, applied analytics
Structured, unstructured data types
National security
Fraud detection
Grand challenge science
Physics, Chemistry, Biology, Weather/climate, energy etc.
Bio/life sciences
Healthcare
Energy/Geophysics
Financial modeling, high frequency and algorithmic trading
Entertainment/media
Auto/aero/mfg.
Consumer
Electronics
Risk informatics: insurance, global, financial, medical etc.
Optimization models
Discovery analytics

Further reading.…..
Currently reviewing

Innovative uses of HPC (LinkedIn.com)

Thank you…..
•Email: sharan dot kalwani at acm dot org

Big Data Everywhere Chicago: High Performance Computing - Contributions Towards Big Data (HPC)

More Related Content

What's hot (20)

Similar to Big Data Everywhere Chicago: High Performance Computing - Contributions Towards Big Data (HPC) (20)

More from BigDataEverywhere (7)

Recently uploaded (20)

Big Data Everywhere Chicago: High Performance Computing - Contributions Towards Big Data (HPC)