Dr. Virendrakumar (Virendra) C. Bhavsar Professor Faculty of Computer Science

1

Dr. Virendrakumar (Virendra) C. Bhavsar

Professor

Faculty of Computer ScienceUniversity of New Brunswick (UNB)

Fredericton, Canada

Supercomputing

2

Outline

• Definitions

Applications

• Hardware

• Software

• Current Status

• University of New Brunswick

• Future

3

Computing

Supercomputing

- A supercomputer is a computer that is at the frontline of current processing capacity, particularly speed of calculation.

High Performance Computing (HPC)/High Productivity Computing

- supercomputing - a subset of HPC

Parallel Computing- many calculations are carried out simultaneously

10**6 Million, 10**9 Billion, 10**12 Trillion

Definitions

4

10**10 Neurons 10**4 Fan-in

- Wires much slower than chips - Millions of times more volume

10**14 Inputs (Connection strngths

10**12 Connection strengths can affect processing in 5 msec

Lower bound on the computational power of brain

~ 10**10 neurons, 10 spikes/sec, 10**14 connections

~10**15 operations/sec or 10**18 bits/sec

Human Brain

5

65K Processors, 5 CM-2 = 1.8 x 10**13 bits/sec

10**5 times slower than brain

Connection Machine CM-2

Early Computers

1950: 5,000 operations/sec; 1970-71: 1 Million Operations/sec

7

1974 - 1 MHz clock1988 – 40 MHz2002 – 2 GHz2009 – P4 3.0 GHz, Quadcore 2.66 MHz

Intel Montecito chip1.72 Billion transistors NVidia 280 series GPU 1.4 Billion transistors

- Circuit complexity doubles every 18 months Computing power at a given cost doubles every 18

months

- Processor clock rates: 40% increase/year + more instr./cycle

- DRAM Access Times: 10% increase/year caches required

Advances in Microprocessor Technology

8

Grand Challenge Applications

- cannot be solved in a reasonable amount of time with today's computers

- Environment, Ecosystems, Molecular engineering, cognition, weapon design, Artificial Intelligence,

(near) Real-Time Applications

- Military/Defense Applications

- Space

-Financial Forecasting; Live data (e.g. online stock market data)

Applications

9

(near) Real-Time Applications

-Google

- Software as a Service (SaaS) delivery model

-ATMs, online banking

Data Intensive Applications

-Walmart – inventory management

- Data Mining

Applications

10

Computational Modeling and Simulation

- Science, Engineering, Social Sciences, …

-Parameter sweep applications

Animation and Movies

Applications

11

Compute Intensive Applications

Massive Data applications

Applications

12

Capability Computing

- Using the maximum computing power to solve a large problem in the shortest amount of time

Capacity computing

- Using efficient cost-effective computing power to solve

- somewhat large problems

- many small problems

Applications

13

Cooling

Speed of Light

Compute Bound Problems I/O Bound problems

Supercomputer Design Challenges

14

Pipelining and Vector Processing

Parallel and Distributed Processing

Liquid Cooling

Non-Uniform Memory Access

Striped Disks (RAID)

Parallel File System

Supercomputer Technologies

15

- Intrinsic parallelism

- Design of parallel algorithms

- Analysis of parallel algorithms

Parallel and Distributed Algorithms

16

PVM and MPI – Loosely connected clusters

OpenMP for Shared Memory Machines

Programming

17

Compilers

Limited success

Automatic Parallelization

Application Checkpointing

18

Roadrunner applications

- National Security

- Planet: Earth and Environmental Sciences

e.g. ground water modeling

- Health: Biology, Chemistry, Life Sciences

- Science: Engineering, Technology

- Universe: Astronomy, Space, Astrophysics

-- Modeling the decay of the US nuclear arsenal

Current Supercomputer

19

Roadrunner

Los Alamos National Laboratory, Los Alamos, NM, USA

- >1 Petaflop (Quadrilion): million billion (10**15) floating-point operations/sec (FLOPS)

-1.71 Petaflop peak

- Weight - 500,000 pounds

- Power - 4 Mega Watt

- Space – 6000 square feet

- Cabling 57 miles

-

Current Supercomputer

20

Roadrunner (Installation Year – 2008)

Los Alamos National Lab, USA

~ 3,250 compute nodes

-Compute Node: Two AMD Opteron dual-core microprocessors

- Each of the Opteron core: Internally attached to one of four enhanced Cell microprocessors.

- Enhanced Cell: double-precision arithmetic faster and can access more memory than can the original Cell in a PlayStation 3. The entire machine will have almost 13,000 Cells and half as many dual-core Opterons.

- Interconnection Network: off-the-shelf Infiniband

Current Supercomputer

21

Roadrunner (Installation Year – 2008)

DOE/NNSA/LANL

System Family - IBM Cluster

System Model - BladeCenter QS22 Cluster

Computer - BladeCenter QS22/LS21 Cluster, PowerXCell 8i 3.2 Ghz / Opteron DC 1.8 GHz , Voltaire Infiniband

Operating System - Linux

Interconnect – Infiniband

Processor - PowerXCell 8i 3200 MHz (12.8 GFlops)

Current Supercomputer

22

Hardware: Building Blocks

• Building blocks – processors, memory, interconnection networks• Processors• Memory – main and secondary storage• Interconnection networks

23

Hardware: Architectures

• Taxonomy: SISD, SIMD, MISD and MIMD• Shared Memory Processing versus Distributed Memory ProcessingSymmetric Multi-Processing (SMP) versus Non-Uniform Memory Access (NUMA) • Processors• Clusters•

24

Special Purpose Supercomputers

• Specially Programmed FPGA chips• Custom VLSI Chips • Reconfigurable Computing • GPUs (Graphics Processing Units)

25

University of

New Brunswick

High Performance Computing and Networking @

University of New Brunswick

“People, Research, Excellence”

ACEnet: Atlantic Computational Excellence Network

Hosting sites:

Member sites:

ACEnet

Atlantic Canada is a distributed environment

$30 million initiative

Waterways make networking solutions difficult (e.g. Cabot Strait)

ACEnet

World-class HPC facilities

Behave as a single, regionally distributed “computational power grid”

Create and operate sophisticated collaboration facilities to bind together geographically dispersed research communities.

Advaced Computational Research Lab (ACRL) Infrastructure

UNB BiologyGary Saunders

UNB ChemistryScott BrownridgeLarry CalhounGhislain DeslongchampsFriedrich Grein

UNB Computer ScienceEric AubanelVirendra BhavsarBrad NickersonRuth Shaw

UNB Text Processing CentreAlan BurkDavid Gants

UNB GeodesyPetr VanícekRichard Langley

UNB MathematicsKeith De’BellAbraham Punnen

UNB Mechanical EngineeringMohammad Bagher AyaniDavid BonhamAndrew GerberMarwan HassanEsam Hussein

UNB PhysicsDr. Eugene K HoDr. Zong-Chao YanDr. Li-Hong Xu

UNB ForestryEvelyn Richards

UNB BiomedicalKevin Englehart

DAL PhysicsAndrew Rutenberg

MTA ChemistryStacey Wetmore

MUN Computer ScienceDwight Kuo

Sick Kids Hospital, TorontoRegis PomesChing-Hsing YuLen Zaifman

StFX Computer ScienceLaurence Yang

UofCalgary Computer SciencePeter TielemanJustin MacCallum

UdeM Environmental StudiesYves Gagnon

UdeM Computer ScienceJalal Almhana

UPEI PhysicsSheldon OppsJames Polson

UofT Computer ScienceHue Sun ChanMaria Sabaye Moghaddam

Major Users

ACEnet at UNB

Fundy: SUN cluster, AMD Opeteron, 632 cores

ACEnet: 3324 cores

Internet connectivity > 2Gbps at UNB

Collaboration Grid

Collaboration gear across Atlantic Canada Lecture rooms equipped so ACEnet sites can share

seminars and participate remotely ACEnet cafés at each site sharing continuous video

feeds Desktop level collaboration equipment for personal

communication

Access Grid streams tens to hundreds of Mbps across the CANARIE network

ACEnet

My Research Work

Special Purpose computers for Military Applications

Design and development of MICRON and PLEXUS

Parallel Monte Carlo Algorithms Graphics and Visualization PaGrid Artificial Intelligence – artificial neural networks, e-

Business Bioinformatics – Canadian Potato Genome project

Future

IBM Cyclops64 – supercomputer on a chip C-DAC initiative for 2010 –petaflop

machine NCSA, USA 2011 petaflop machine NASA, SGI and Intel Pleiades – 10

petaflop by 2012 1 Exaflop (10**18 flops) by 2019 Human brain neural simulations – 10

exaflop by 2025 2-week Full Weather modeling – 1 zeta

flops (10**21 flops) by 2030