Top Banner
BLUE GENE/L Sapnah Aligeti CMPS 5433
34
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Blue Gene

BLUE GENE/LSapnah AligetiCMPS 5433

Page 2: Blue Gene

Outline• History about supercomputers• Manufacturers / Partners of Blue Gene/L• Why was it created?• Who are the customers?• How much does it cost?• Processors / Memory / Scalability• Stepwise Structure• Hardware Architecture• Interconnection Network• Software• Advantages• Applications

Page 3: Blue Gene

A LITTLE ABOUT SUPERCOMPUTERS……

• IBM’s Naval Ordnance Research Calculator.

• IBM's Blue Gene/L.

Page 4: Blue Gene

……A LITTLE ABOUT SUPERCOMPUTERS (CONTD)

360000000000000

floating-point

operations per

second (TFLOPS) in March, 2005.

15,000 operations per second.

Page 5: Blue Gene

……A LITTLE ABOUT SUPERCOMPUTERS (CONTD)

Page 6: Blue Gene

MANUFACTURER / PARTNERS

• 1999 - 100M $ PROJECT BY IBM FOR THE US DEPT OF ENERGY (DOE) - BLUE GENE/L - BLUE GENE/C (CYCLOPS) - BLUE GENE/P (PETAFLOPS)

• 2001 - PARTNERSHIP WITH LAWRENCE LIVEMORE NATIONAL LABORATORY (FIRST CUSTOMER)

Page 7: Blue Gene

TWO MAIN GOALS OF BLUE GENE/L

• to build a new family of supercomputers optimized for bandwidth, scalability and

the ability to handle large amounts of data while consuming a fraction of the

power and floor space required by today's fastest systems.

• to analyze scientific and biological problems

(protein folding).

Page 8: Blue Gene

CUSTOMERS• 64 rack machine to Lawrence Livermore National Laboratory,

California

• 23 Feb 2004 – 6 rack machine to ASTRON, a leading astronomy organization in the Netherlands to use IBM's Blue Gene/L supercomputer technology as the basis to develop a new type of radio telescope capable of looking back billions of years in time.

• May/June 2004 – 1 rack system to Argonne National Laboratory, Illinois

• Sept 2004 IBM - 4 rack Blue Gene/L supercomputer to Japan's National Institute of Advanced Industrial Science and Technology (AIST) to investigate the shapes of proteins.

• 6 Jun 2005 - 4 rack machine to The Ecole Polytechnique Federale de Lausanne (EPFL), in Lausanne, Switzerland to simulate the workings of the human brain .

Page 9: Blue Gene

COST• The initial cost was 1.5 M $/rack

• The current cost is 2M $/rack

• March 2005 – IBM started renting the machine for about $10,000 per week to use one-eighth of a Blue Gene/L rack.

Page 10: Blue Gene

PROCESSORS / MEMORY / SCALABILITY

PROCESSOR• 65,536 DUAL PROCESSOR NODES.• 700 MHZ POWER PC 440 PROCESSOR.

MEMORY• 512 MB of dynamic random access

memory (DRAM) per node.

SCALABILITY• BLUE GENE/L IS JUST THE FIRST

STEP………

Page 11: Blue Gene

THE BLUE GENE/LTHE STEPWISE STRUCTURE……

Page 12: Blue Gene

THE BLUE GENE/L

Page 13: Blue Gene

THE BLUE GENE/LTHE RACK/CABINET

Page 14: Blue Gene

THE BLUE GENE/LTHE NODE CARD

Page 15: Blue Gene

THE BLUE GENE/LTHE COMPUTE CARD

Page 16: Blue Gene

THE BLUE GENE/LTHE CHIP

Page 17: Blue Gene

THE BLUE GENE/L

Page 18: Blue Gene

THE BLUE GENE/L

Page 19: Blue Gene

THE BLUE GENE/L

Page 20: Blue Gene

BLUE GENE/L I/O ARCHITECTURE

Architecture: Top view:

Page 21: Blue Gene

HARDWARE

• 65,356 Compute nodes– ASIC (Application-Specific Integrated

Circuit) – ASIC includes two 32-bit PowerPC

440 processing cores, each with two 64-bit FPUs (Floating-Point Units)

– compute nodes strictly handle computations

• 1024 i/o nodes – manages communications for a

group of 64 compute nodes.

• 5 Network connections

Page 22: Blue Gene

Interconnection Network

• 3D Torus

• Global tree

• Global interrupts

• Ethernet

• Control

Page 23: Blue Gene

3D TORUS n/w FOR 64 NODES (4 * 4 * 4)

• http://hpc.csie.thu.edu.tw/docs/Tutorial.pdf

Page 24: Blue Gene

Torus n/w (contd)

• Primary connection• Torus n/w connects all the 65,536

compute nodes (32 * 32 * 64).• One node connects to 6 other nodes.• Chosen because provides high

bandwidth nearest neighbor connectivity

• Single node consists of single ASIC and memory.

• Dynamic adaptive routing.

Page 25: Blue Gene

SOFTWARE

• The main parallel programming model for BG/L is message passing using MPI (message passing interface) in C, C++, or FORTRAN.

• Supports global address space programming models such as Co-Array FORTRAN (CAF) and Unified Parallel C (UPC).

• The I/O and external front-end nodes run Linux, and the compute nodes run a kernel that is inspired by Linux.

Page 26: Blue Gene

Advantages

• Scalable

• Less space (half of the tennis court)

• Heat problems most supercomputers face

• Speed

Page 27: Blue Gene

Limitations

– Memory Limitation (512 MB/node)

– Simple node kernel (does not support forks, threads)

Page 28: Blue Gene

Applications

• BLUE BRAIN PROJECT, 6 JUNE IBM and Ecole Polytechnique Fédérale

de Lausanne (EPFL), in Switzerland to study the behavior of the brain and model it.

• PROTEIN FOLDINGAlzheimer’s disease

Page 29: Blue Gene

Future developments????

• Article published in “THE STANDARD”, china’s business newspaper dated May 29

– Military hopes such a development will allow pilots to control jets using their mind

– Allow wheelchair users to walk

Page 30: Blue Gene

References

• IBM, Journal of Research and Development, volume 49, November 2005.

• Goolge News.• http://www.linuxworld.com/read/48131.htm• http://sc-2002.org/paperpdfs/pap.pap207.pdf• http://www.ipab.org/Presentation/sem04/04-02-

2.pdf• http://www.desy.de/dvsem/WS0405/

steinmacherBurow-20050221.pdf• www.scd.ucar.edu/info/UserForum/presentations/

loft.ppt

Page 31: Blue Gene
Page 32: Blue Gene

ASIC

Page 33: Blue Gene

GENERAL CONNECTION

Page 34: Blue Gene

What is a kernel??• In computer science, the kernel is the

fundamental part of an operating system. It is a piece of software responsible for providing secure access to the machine's hardware to various computer programs. Since there are many programs, and access to the hardware is limited, the kernel is also responsible for deciding when and how long a program should be able to make use of a piece of hardware, which is called multiplexing. Accessing the hardware directly can be very complex, so kernels usually implement some hardware abstractions to hide complexity and provide a clean and uniform interface to the underlying hardware, which helps application programmers.