+ All Categories
Home > Documents > From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o...

From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o...

Date post: 21-Aug-2021
Category:
Upload: others
View: 2 times
Download: 0 times
Share this document with a friend
31
Robert C. Bell – CSIRO and Bureau of Meteorology From CSIRAC to Cray (again) and onwards CSIRO INFORMATION MANAGEMENT & TECHNOLOGY SCIENTIFIC COMPUTING
Transcript
Page 1: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

• Robert C. Bell – CSIRO and Bureau of Meteorology

From CSIRAC to Cray (again) and onwards

CSIRO INFORMATION MANAGEMENT & TECHNOLOGY

SCIENTIFIC COMPUTING

Page 2: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

•The past: •history of computing support for science in

Australia •The present: • the Bureau’s Cray system •The future: •computing for science

Robert C. Bell

Summary

2 |

Page 3: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Robert C. Bell

Dinosaurs

3 |

Page 4: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Robert C. Bell

4 |

Page 5: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

History of Compute performance in CSIRO

0.0000001

0.000001

0.00001

0.0001

0.001

0.01

0.1

1

10

100

1000

10000

100000

1000000

10000000

1940 1960 1980 2000 2020

Gfl

op

/s

CSIRO Peak Computing Systems

Peak Speed(Gflop/s)

CSIRAC

CDC

CDC Cyber

CDC Cyber

Cray

NEC vector

raijin & magnus

APAC SGI

NCI Sun

Robert C. Bell 5 |

Page 6: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

History of Computing in the Bureau

0.0001

0.001

0.01

0.1

1

10

100

1000

10000

100000

1000000

10000000

1965 1970 1975 1980 1985 1990 1995 2000 2005 2010 2015 2020

Peak P

erfo

rm

an

ce (

flo

ps/

sec)

Years

Peak Performance

IBM 360/65

FACOM/Fujistu M200 CDC/ETA 10P

Cray X-MP, Y-MP

NEC SX-4

NEC SX-5

NEC SX-6

Sun Constellation

Oracle

Cray XC40

Cray XC+

1 MegaFlop

1 PetaFlop

1 GigaFlop

1 TeraFlop

10 MF

100 MF

10 GF

100 GF

10 TF

100 TF

10 PF

Robert C. Bell 6 |

Page 7: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Early computing

Robert C. Bell

Page 8: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

• CSIRAC in CSIR Radiophysics from 1949 •Moved to University of Melbourne in 1956 •Dick Jensen ran first numerical forecasts in

Australia on CSIRAC • CSIRO Computing Research – CDC 3600 (used by

Bureau until IBM 360/65 systems in 1968) • CSIRO: Csironet: CDC Cybers, then post-Csironet

Crays • Bureau: FACOMs, ETA, Crays •HPCCC: NEC

Robert C. Bell

Early history

8 |

Page 9: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

9 | Robert C. Bell

Page 10: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

CSIRO HPC Systems

Robert C. Bell 10 |

Page 11: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Robert C. Bell

11 |

Page 12: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Provision of: •Systems – compute •Storage – HSM •Software – Open Source treasury •Services – Help Desk, backups, etc •Support – people with knowledge

Robert C. Bell

History

12 |

Page 13: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

• New data centre • Cray XC40 – widely adopted in the community • Features: • Intel Xeon Haswell processors: • FMA3: d=round(a x b + c) ; Variable clock speed

• Dense packaging: 384 processors per cabinet: 90 kW • Special compute-node Linux version: low jitter • Water cooled • Two partitions: for resiliency: compute and storage • Aries interconnect: • 48-port routers, 500 Gbyte/s; dragonfly topology

Robert C. Bell

The present: Bureau’s Cray system

13 |

Page 14: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Cray XC Rank1 Network

o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Robert C. Bell 14

Page 15: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Australis HPC

Parameter Oracle HPC

System 2015 System 2018 System

Relative Increase

Processor

Intel Xeon Sandy Bridge

6-core 2.5GHz

Intel Xeon Haswell E5-2690v3

12-core 2.6GHz

Intel Xeon

Haswell + Skylake

Increase relative to Oracle HPC

System Ngamai

Nodes 576 2,160 4,112 2015: 3.8x 2018: 7.1x

Cores 6,912 51,840 129,920 2015: 7.5x

2018: 18.8x

Aggregate

Memory 36.9 TB 276 TB 651 TB

2015: 7.5x

2018: 17.7x

Global Filesystem Technology

Oracle Lustre 1.8.8

Sonexion Lustre 2.5.1+

Sonexion Lustre 2.x

Usable Storage 214 TB 4,320 TB 8,640 TB 2015: 20.3x 2018: 60.5x

Storage

Bandwidth 16 GB/s 135 GB/s 306 GB/s

Compute

Interconnect

Mellanox Infiniband

QDR 40Gb/s

Cray Aries

93 – 157Gb/s

Cray Aries

93 – 157Gb/s

Typical Power

Use (kiloWatts) 200 kW 865 kW 1,648 kW

2015: 4.3x

2018: 8.2x

Sustain system

performance (SSP)

16 253 618 2015: 15.6x

2018: 38.1x

Top 500 Rmax Linpack (TF)

104 1663 5000+ 2015: 16.0x 2018: 50.0x

Robert C. Bell 15 |

Page 16: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Operational Capacity (Half of total capacity)

1.E+04

1.E+05

1.E+06

1.E+07

1.E+08

2010 2011 2012 2013 2014 2015 2016 2017

Syste

m P

erfo

rm

an

ce (

Top

50

0 H

PL R

max)

NWP Computing Capacity of National Meteorological Centres

100T

PetaFlops

10P

100P

10T

ECMWF (1.8PF)

UK Met Office

Korea Met Agency

Environment Canada

Australian Bureau of Meteorology

#1 Machine 54 Pflops China

Bureau of Meteorology (N/A)

100T

PetaFlops

10P

100P

10T

US-NWS

#500 Trend Line

#1 Trend Line US-NWS (2.5PF)

UK Met Office (2.8PF)

ECMWF

Robert C. Bell 16 |

Page 17: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

•Systems – no more easy gains? •Storage – cloud? •Software – narrowing of options because of

architectures •Services – computer science, collaboration,

visualisation •Support – people with knowledge

Robert C. Bell

Future possibilities

17 |

Page 18: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Robert C. Bell

Page 19: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

Stream2 - 2004

Robert C. Bell

Page 20: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

•Systems – more cores, not much faster? •Limited bandwidth gains •Accelerators: for some applications •CSIRO accelerated computing program –

GPU systems – some success •Commodity market – ARM processors?

Robert C. Bell

Future systems

20 |

Page 21: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

• Ten years ago, I wrote: • Users typically want every file kept and backed-up, and would be

happy to use only one file system, globally visible across all the systems they use, with high-performance everywhere, and infinite capacity!

• A user added that they want all of the above, at zero cost!

• Storage – cloud? Regulatory, cost and protection issues • Storage for HPC – Lustre, fast, parallel, but not very

reliable. • Object stores rather than filesystems? • Tape persists, even in cloud • CSIRO Cloud storage

Robert C. Bell

Future Storage

21 |

Page 22: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

CSIRO area

22 | Robert C. Bell

Page 23: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

23 | Robert C. Bell

Page 24: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

•Explosion in consumer software but, in HPC, narrowing of options because of architectures

•Massive investment in major models •Collaboration essential •CSIRO eResearch model

Robert C. Bell

Future: Software

24 |

Page 25: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

E-enablement

equipping the organisation to perform activities more

efficiently through electronic means. The e-enablement

layer includes general computing and storage,

networks and office productivity tools both on-

premise and remote

Consolidate infrastructure

Increase reliability/availability

Achieve economies of scale

Virtualise

eResearch

the rapid evolution of research methodologies enabled by

information technologies and tools. The eResearch layer

includes scientific computing and storage, research data

management, visualisation, advanced collaboration

technologies and productivity tools.

Expand scientific computing

Embed enterprise-level research data management

Extend collaboration/e-tools

Develop visualisation services

Robert C. Bell

Page 26: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

IM&T eResearch Supports the Entire Science Data Workflow

Workflow Services

Research Data Services

Information Services Research

Question

Research Design

Data Collection

Processing & Analysis

Data & Workflow Archiving

Disseminate & Publish

Data Re-Use

Measure Impact

IM&T eEnablement Services: High Speed Networks, Application Development, Tele-presence, Office

Productivity, Collaboration Tools

Scientific Computing & Visualisation

Outreach & eResearch Planning

Advanced Collaboration

Robert C. Bell 26 |

Page 27: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

• Submission History

eResearch Project Services

Robert C. Bell

0

10

20

30

40

50

60

70

80

90

100

RFP Responses by Type

Workflow

Sustainability

Data

Vizualisation

Compute

Unique

Page 28: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

•File protection (backup) •Computer science •Collaboration •Visualisation •Buzz words: e.g. Big Data

Robert C. Bell

Future: services

28 |

Page 29: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

•People with knowledge •Critical mass •Training – little HPC knowledge from

undergraduates (Fortran?) •Staff with commitment to science support •Bureau: Scientific Computing Services

Robert C. Bell

Future: support

29 |

Page 30: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

• We’ve come a long way in computing for atmospheric and ocean sciences

• Major acquisition by Bureau for operational computing

• Hard to see future, but we know some dead-ends

• Forecast models running on phones?

Conclusion: From CSIRAC to Cray (again) and onwards

Page 31: From CSIRAC to Cray (again) and onwards · Cray XC Rank1 Network o Chassis with 16 compute blades o 128 Sockets o Inter-Aries communication over backplane o Per-Packet adaptive Routing

CSIRO IM&T Scientific Computing Services Robert C. Bell

t +61 3 9669 8102 | +61 3 9545 2368

e [email protected]

w www.hpsc.csiro.au

CSIRO IM&T SCIENTIFIC COMPUTING SERVICES

Thank you


Recommended