47
1 © Bull, 2012 The views expressed are those of the author and do not reflect the official policy or position of Bull 2012/11/19 Xavier Vigouroux Business Development Manager igh erformance omputing The views expressed are those of the author and do not reflect the official policy or position of Bull

igh erformance omputing - ens-lyon.fr

  • Upload
    others

  • View
    8

  • Download
    0

Embed Size (px)

Citation preview

Page 1: igh erformance omputing - ens-lyon.fr

1 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l

2012/11/19 Xavier Vigouroux

Business Development Manager

igh erformance omputing

The views expressed are those of the author and do not reflect the official policy or position of Bull

Page 2: igh erformance omputing - ens-lyon.fr

2 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 3: igh erformance omputing - ens-lyon.fr

3 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Yes ! It’s worth to invest

Page 4: igh erformance omputing - ens-lyon.fr

4 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 5: igh erformance omputing - ens-lyon.fr

5 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Who am I

19

96

Ph

D

Page 6: igh erformance omputing - ens-lyon.fr

6 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 7: igh erformance omputing - ens-lyon.fr

7 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l

7 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Bull: European leader in mission-critical digital systems

recognized worldwide

in secure systems

OPERATING IN 50 COUNTRIES

€1.3bn REVENUES

+29% growth in profitability in 1st quarter 2012

+4,6% growth in 2011

+23% Efforts in research in 2011

Page 8: igh erformance omputing - ens-lyon.fr

8 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 9: igh erformance omputing - ens-lyon.fr

9 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l

m.

Linpack

Ax = b Flops/s

top500.

How large is a machine ?

Page 10: igh erformance omputing - ens-lyon.fr

10 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Top 3

10 051 Tflops/s

11 280 Tflops/s

705 024 sparc cores

16 325 Tflops/s

20 132 Tflops/s

1 572 864 PowerPC cores

17 590 Tflops/s

27 112 Tflops/s

299 008 cores AMD

261 632 cores GPU

Page 11: igh erformance omputing - ens-lyon.fr

11 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l TOP500

1,E+00

1,E+01

1,E+02

1,E+03

1,E+04

1,E+05

1,E+06

1,E+07

1,E+08

1,E+09

0 10 20 30 40 50 60

Rpeak Rmax

Page 12: igh erformance omputing - ens-lyon.fr

12 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

TERA 100 1.25 PetaFlops

140 000+ Xeon cores

256 TB memory

30 PB disk storage

500 GB/s IO throughput

580 m² footprint

CURIE 2 PetaFlops

90 000+ Xeon cores 148 000 GPU cores

360 TB memory

10 PB disk storage

250 GB/s IO throughput

200 m² footprint

IFERC 1,5 PetaFlops

70 000+ Xeon cores

280 TB memory

15 PB disk storage

120 GB/s IO throughput

200 m² footprint

3 petaflop-scale systems

Page 13: igh erformance omputing - ens-lyon.fr

13 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

TERA 100 GPU-based extension 198 bullx B505 accelerator

blades 396 NVIDIA® Tesla™

M2090 GPU processors 202,752 GPU cores

CURIE GPU-based extension 144 bullx B505 accelerator

blades 288 NVIDIA® Tesla™

M2090 GPU processors 147,456 GPU cores

BSC GPU-based system 126 bullx B505 accelerator

blades 252 NVIDIA® Tesla™

M2090 GPU processors 129,024 GPU cores

3 large GPU based systems

#7 – green500

Page 14: igh erformance omputing - ens-lyon.fr

14 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Source: http://www.top500.org – list nov’12

Page 15: igh erformance omputing - ens-lyon.fr

15 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 16: igh erformance omputing - ens-lyon.fr

16 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Load the supercomputer

Page 17: igh erformance omputing - ens-lyon.fr

17 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Ramp up the users

Tier 0

• 1 Petaflop/s – 50000 cores

• European, PRACE

• TGCC, IFERC

Tier 1

• 100 Teraflop/s – 5000 cores

• National

Meso Centres

• 10 Teraflop/s – 500 cores

• Regional

Page 18: igh erformance omputing - ens-lyon.fr

18 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Cooling

30°C – 80KW

6°C – 30KW

12kW/rack

Page 19: igh erformance omputing - ens-lyon.fr

19 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Cooling

Water has closed as possible to the heat source

Water can be hotter (as delta T is key)

Room can be hotter (remove CRAC)

But, maintainability is key ! No change in maintenance process

CPU can be changed,

DIMM can be changed

Blades can be removed

30°C – 80KW

Page 20: igh erformance omputing - ens-lyon.fr

20 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Power Feeding

82

83

84

85

86

87

88

89

90

91

92

93

0 50 100 150 200 250 300%

Load (V)

PSU Efficiency (240V)

15% 10%

Page 21: igh erformance omputing - ens-lyon.fr

21 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Efficient Hardware

2 x CPUs 2x GPUs/Xeon Phis

Page 22: igh erformance omputing - ens-lyon.fr

22 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Bull - Energy aware batch scheduler

Fix the CPU frequency

Effective CPU Frequency Job Power consumption

150

200

250

300

350

400

450

0 1000000 2000000 3000000

J/s

freq

15000

17000

19000

21000

23000

25000

0 50 100

Co

nsum

ed E

nerg

y (J

)

Duration (s)

$# srun --cpu-freq=2700000 --resv-ports -N2 -n64 ./cg.C.64& $# sacct -j 58 -format=jobid,elapsed,aveCPUFreq,consumedenergy

JobID Elapsed AveCPUFreq ConsumedEnergy ------------ ---------- ---------- -------------- 66 00:00:49 2640340 19668

Page 23: igh erformance omputing - ens-lyon.fr

23 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Use efficient Library

Courtesyof Jack DONGARRA – MAGAM + starPUi

Page 24: igh erformance omputing - ens-lyon.fr

24 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 25: igh erformance omputing - ens-lyon.fr

25 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l 2 Key elements

Computing the function

Managing Data

function Data Results

Page 26: igh erformance omputing - ens-lyon.fr

26 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Data Movement and Floating point

FPS-164 and VAX (1976)

11 Mflop/s; transfer rate 44 MB/s

Ratio of flops to bytes of data movement: 1 flop per 4 bytes transferred

Nvidia Fermi and PCI-X to host

500 Gflop/s; transfer rate 8 GB/s

Ratio of flops to bytes of data movement: 62 flops per 1 byte transferred

Flop/s are cheap, so are provisioned in excess

Page 27: igh erformance omputing - ens-lyon.fr

27 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [DATA] Non Volatile RAM

Source: ITRS

Page 28: igh erformance omputing - ens-lyon.fr

28 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [DATA] 3D Wioming from CEA leti

Wide IO Memory Interface Next Generation

Courtesy of CEA leti

Page 29: igh erformance omputing - ens-lyon.fr

29 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [DATA] Hybrid Memory Cube

Wide I/O standard comes from mobile world

True « 3D » (with Through Silicon Vias »)

Source: http://www.synopsys.com, http://blogs.intel.com/research/2011/09/15/hmc/,

http://www.semiconwest.org/sites/semiconwest.org/files/docs/Shekhar%20Borkar_Intel.pdf

Bandwidth Energy

DDR-3 10,66 GB/s 50-70pJ/bit

HMC 128GB/s 8 pJ/bit

Page 30: igh erformance omputing - ens-lyon.fr

30 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [DATA] Photonics

Photonic is getting closer to the CPU

Basic blocks are in place:

Modulator

Multiplexer

Routing

Receiver

Off chip Energy cost very close to On chip Energy cost

To go deeper: “Silicon Photonics for Exascale Computing” Keren Bergman, Columbia universisty Lightwave Research Laboratory

Page 31: igh erformance omputing - ens-lyon.fr

31 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [FUNCTION] Near threshold Voltage

http://www.realworldtech.com/near-threshold-voltage/

Drawback: surface is increasing

Page 32: igh erformance omputing - ens-lyon.fr

32 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [FUNCTION] CPU frequency

Source: top500 - mean freq

0

500

1000

1500

2000

2500

3000

19

84

19

85

19

86

19

87

19

88

19

89

19

90

19

91

19

92

19

93

19

94

19

95

19

96

19

97

19

98

19

99

20

00

20

01

20

02

20

03

20

04

20

05

20

06

20

07

20

08

20

09

20

10

20

11

20

12

Freq

uenc

y (

MH

z)

?

?

Page 33: igh erformance omputing - ens-lyon.fr

33 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l [FUNCTION] Mask reduction

1

10

100

1000

10000

1960 1970 1980 1990 2000 2010 2020 2030

Tech

nolo

gy

nod

e si

ze (

nm)

Page 34: igh erformance omputing - ens-lyon.fr

34 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 35: igh erformance omputing - ens-lyon.fr

35 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l TOP500

1,E+00

1,E+01

1,E+02

1,E+03

1,E+04

1,E+05

1,E+06

1,E+07

1,E+08

1,E+09

0 10 20 30 40 50 60

Rpeak Extrap Peak Rmax Extrap Max

Page 36: igh erformance omputing - ens-lyon.fr

36 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Is it a long way to Exascale

Jun ‘97 Nov ‘ 08 Nov ‘12 Jun ’19 asci red roadrunner Titan Eflop

Sandia ORNL ORNL ???

MW 0,50 2,48 8,3 20

Eflop max 1,45E-06 0,001 0,027 1

nodes 7264 6480 18688 64000

racks 104 296 200 300

U/rack 30 42 42 42

Concur. 9152 129’600 50 M O(109) Year factor Year factor Year factor

Gflop/W 2,91E-03 1,549 0,4 1,651 3,3 1,519 50,0

Tflop/node 2,00E-04 1,798 0,2 1,708 1,5 1,441 15,6

W/node 68,8 1,161 383,3 1,035 439,3 0,949 312,5

nodes/rack 69,8 0,904 21,9 1,437 93,4 1,135 213,3

W/rack 4807,7 1,050 8390,1 1,487 41045,0 1,077 66666,7

Tflop/U 4,66E-04 1,579 0,09 2,455 3,23 1,637 79,37

W/U 160,3 1,019 199,8 1,487 977,3 1,077 1587,3

Page 37: igh erformance omputing - ens-lyon.fr

37 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Node evolution : « à la Phi »

60 cores @22nm => ~500 cores @8nm

1 Tflops @22nm => ~8 Tflops @ 8nm

With µarch improvement, it should be less !

Phi power consumption will have to be around 150W-300W

Between 200 and 400 Phi per Rack

1

10

100

1000

10000

1960 1980 2000 2020 2040

Tech

nolo

gy

nod

e si

ze (

nm) Gflop/W 50,0

Tflop/node 15,6 1-2 Phi

W/node 312,5

nodes/rack 213,3

Tflop/U 79,37 10 Phi

W/U 1587,3

Page 38: igh erformance omputing - ens-lyon.fr

38 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

[#freq] x [#flop/cycle] x [#cores]

1050 MHz – 700 MHz 16 - 64

22’000’000

Page 39: igh erformance omputing - ens-lyon.fr

39 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Gene Amdahl John Gustafson

Page 40: igh erformance omputing - ens-lyon.fr

40 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Stack

Data Center

Supercomputer

Runtime

Hurry UP

Speed UP

Keep the Pace

Libraries

Application

Page 41: igh erformance omputing - ens-lyon.fr

41 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Strong Integration

Ressource Manager

Data exchange

Library

Code

Resiliency

Topology

See Fastforward approach

See MAGMA [UTK] with StarPU [INRIA]

Page 42: igh erformance omputing - ens-lyon.fr

42 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Supercomputer Suite

All functionalities o Cluster management o Development factory o Execution environment o Data storage and access All sizes o From department o to Top 10

Installed, deployed and operated as a

single software

Modular: Get what you need

when you need

Best of breed o Linux, o OpenMPI, o HPC Toolkit, o Nagios, o OFED, o Slurm, o Lustre, o Shine, ... + Bull added value

COMPLETE FLEXIBLE

INTEGRATED

OPEN

Page 43: igh erformance omputing - ens-lyon.fr

43 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l Applications and Performances team

©

B

ull

One Bull expert (PhD) for each segment

Reservoir Simulation Rendering / Ray Tracing

Climate / Weather Ocean Simulation

Structural Mechanics Implicit

Structural Mechanics Explicit

Computational Fluid Dynamics

Electro-Magnetics Computational Chemistry

Quantum Mechanics

Data Analytics

Computational Chemistry Molecular Dynamics

Computational Biology

Seismic Processing

Page 44: igh erformance omputing - ens-lyon.fr

44 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l

44 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l

Page 45: igh erformance omputing - ens-lyon.fr

45 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 46: igh erformance omputing - ens-lyon.fr

46 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l T

he v

iew

s e

xpre

ssed a

re t

hose o

f th

e a

uth

or

and d

o n

ot

refle

ct th

e o

ffic

ial polic

y o

r positio

n o

f B

ull

Page 47: igh erformance omputing - ens-lyon.fr

47 © Bull, 2012

The

vie

ws

exp

ress

ed

are

tho

se o

f the

aut

hor

and

do

no

t re

flect

the

offi

cia

l po

licy

or

po

sitio

n o

f Bul

l

For any question [email protected]