hpc & big data @ atos...curie - 2011 1st prace petascale supercomputer intel e5 “early ird”...
TRANSCRIPT
© Atos 2015
HPC & Big Data @ Atos
Damien Déclat, 27 Nov 2015
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Atos is a leader in digital services
delivering Systems Integration services,
Consulting, Managed Services & BPO, Cloud
operations, Big Data & Cyber-security solutions,
as well as transactional services. Atos is
focused on business technology that powers
progress and helps organizations to
create their firm of the future.
Atos operates under the brands Atos,
Atos Consulting, Atos Worldgrid, Bull,
Canopy, and Worldline.
Atos is a Societas europaea (SE)
Atos is the Worldwide Information
Technology Partner for the Olympic &
Paralympic Games and is listed on the
Euronext Paris market.
Profile
billion pro formaannual revenue 2014
business technologists
countries around the world
02
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
liber
mobull
ied protection
hoox
extreme factory
bullionbullx
Through its Bull technologies, Atos develops the highperformance computing platforms, security solutions, sowareappliances and services allowing its customers to monetize and protect their information assets.
Bull, Atos technologies:For extreme performance and extreme security
Cyber security & defense
Scientific computing
ITmodernization
Business computing
escala
6
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
HPC today… Top500 trend for the last 10 years
07
23 24 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46
• 22 systems in Top500 of Nov 2015
• #1 in computing power installed in Europe this year (6 PFlops installed)
in
▶ Already more than
10 PFlops coming up
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Proven know-how for large productive systems
8
CURIE - 20111st PRACE Petascale supercomputerIntel E5 “Early Bird”150 GB/s Lustre2 PFlops peak
OCCIGEN 2014TIER0 Supercomputer, CINESDLC technology2.1 PFlops peak 250 + 300GB/s – Lustre & DMF
TAURUS 2013-20141st BULL PetaFlops Supercomputer in Germany1 PFlops peakLustre
DKRZ 2014-2016ClimatologyDLC technology3 PFlops45 PB @ 480 GB/sLustre + HPSS
CARTESIUS 2013-20141st Bull Petascale Supercomputer in Netherland DLC technology1.3 PFlops8PB @ 220 GB/s - Lustre
HELIOS 2011-2014ITER Community1.7PFlops peakX86 + PHI+100GB/s – Lustre
BEAUFIX PROLIX 2013-20141st Intel E5 v3 supercomputer in production wwDLC technology1 PFlops peakExtension to 5 PFlops in 2016
Santos Dumont 2015Largest supercomputer in LATAMDLC Technology1 PFlops peakMobull
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
The new HPC paradigm
9
From technology… … to usage
1970 2010 2020
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Reaching better Productivity and higher Efficiency
10
Efficiency• Optimization of applications
• Better knowledge of technology
• Solution Design
Productivity• Smart computing
• Operational Excellence
• Design & develop with future in mind
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Sequana, live innovation to the fullest
All-in-one sequana cell
▶ 2 compute cabinets
▶ 1 server cabinet including L1&L2 interconnect switches and a management server
▶ 288 compute nodes- Intel Xeon Broadwell-EP, Xeon PHI KNL,
Skylake-EP- NVIDIA GPU Pascal accelerator
▶ 1 interconnect- Infiniband EDR, BXI
▶ Open and multi-technology
– to preserve investments over the long-term
– to fit with all kinds of usages
▶ Ultra dense and scalable
– Large building blocks to facilitate scaling
– Embedding the fastest interconnects
▶ Ultra-energy efficient
▶ Easy administration
– Pre-installed cell
– Integrated service nodes
▶ 1 Petaflop in 1 sequana cell (2016)
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Sequana, live innovation to the fullest
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Sequana, live innovation to the fullest
Compute Rack “base”
90kW PSUs
25GB/s sideplane cables (NIC L1 )
x3 Hydraulics chassis (heat exchange)
x48 Compute blades
Switch rack
12kW PSUs
25GB/s copper cable midplane (L1 L2)
Ethernet Top/Leaf switches for island mgt. (behind midplane)
Management servers
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Bull S6000, for compute and data intensive applications
Up to 16 sockets
Up to 24 TB memory
300+ systems
already deployed
Uni WarwickUK
High performance, high-volume data
analysis
First of its kind in the UK
Open-architecture multi-platform
analytical capability
LNCCBrazil
the largest in-memory general
purpose system in latin America
Chemistrymolecular dynamics
life scienceClimatology
Uni Rijeka
Croatia
Multi-TB in-memory analytics
MolecularDynamics
CNAGSpain
Accelerate fundamental research in Encology
ultra fast end-to-end genome
decoding platform
CalmipFrance
The first Haswel EX deployed
worldwide
On going collaboration around key
applications for Calmip
Corporate Presentation
CEPP Adopters
Customer
Challenges
▶ Technology is moving fast
▶ Even on the best hardware, Applications must be adapted
▶ Applications are key for productivity
▶ Anticipation & integration of technology trends
Center for Excellence in Parallel ProgrammingCenter for Exploration and Innovation
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
The Fast Start ProgramAccelerating Adoption and Maximizing ROI
Start of production
Time
Sup
erco
mp
ute
ru
tilil
izat
ion
rate
Ad
op
tio
n c
urv
e
without fast start
with fast start
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Centre d’Excellence en Programmation Parallèle
Code
Targeted time
Targeted #coresWorkshop in
Grenoble Oct 2014
Bull internalmeeting
Nov 2014
Customermeeting Dec 2014
Customermeeting Jan 2015
T0 S1 (in s) P0 S1 ∆ ∆ ∆ ∆
ICON_APEASIS 599 4 509 -12 % -12 % -12 % -8 %
ICON_APEOPTIM 599 3 381 -53 % -40 % -34 % -31 %
ICON_LAMASIS 22,6 16 000 NA +13 % +14 % +14 %
ICON_LAMOPTIM 18,0 16 000 NA -2 % -2 % -2 %
CCLMASIS 324 752 -4 % -4 % -1 % +1 %
CCLMOPTIM 324 662 -19 % -19 % -8 % -6 %
FESOMASIS 49,9 846 +6 % +6 % +8 % +2 %
FESOMOPTIM 49,9 786 0 % 0 % 0 % +5 %
EMACASIS 13 675 259 -8 % -6 % +7 % +7 %
EMACOPTIM 13 811 184 NA NA -7 % +6 %
MPI-ESMOPTIM 10 961 163 -12 % -9 % +4 % +4 %
METRASOPTIM 4 400 10 NA NA +25 % +25 %
EH6-CDI-PIO(r) 1 555 768 NA +25 % +33 % +33 %
EH6-CDI-PIO(p) 1 442 744 NA +20 % +24 % +24 %
With Turbo
Without Turbo
lack of performance > 5% lack of performance < 5% performance achieved
+12%
+7%
+29%
∆ : difference relative
to commitments
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Incorporating the most recent technological advances to boost academic research
18
The Scientific Grouping CALMIP, which stands for Computation in Midi-Pyrénées, was founded in 1994 by 17 Research Laboratories of the city of Toulouse and the province of Midi-Pyrénées. Its objective is to promote the use of new technologies in Scientific Computation in the researcher community. Through scientific events CALMIP encourages interdisciplinary exchanges and creates a dynamic plat-form concerning the use of HPC at the regional scale.
▶ Replace the existing supercomputer so as to continue to offer to the CALMIP research community a system incorporating the most recent technological advances
▶ A production system, able to serve the scientific community for the four years to come.
Business challenge
▶ Direct liquid cooling
▶ 612 bullx DLC B710 compute nodes, each equipped with 2 Intel®
Xeon® E5 v2 processors (12.240 cores in total)
▶ A system with a peak performance of 274 Tflops, for an electrical consumption of 244 kW (LINPACK).
Solution
▶ A system providing a peak performance increased by a factor of 7 compared to the previous system, while the energy consumption is only doubled.
▶ The system is installed in a new building shared with Météo France and their own bullx supercomputer based on the same technology.
Benefit
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Bringing weather forecasting to the next level at Météo-France
Météo France is France’s national weather forecast service.
Its core mission is to issue warnings in case of extreme weather events. This implies that it operates 24 hours a day and every day of the year.
Météo France also conducts research on climate change.
▶ Production must not be disrupted▶ Replace current 41.8 Tflops vector system by a scalar
supercomputer ▶ Issue : porting and optimizing the codes for the new configuration
Business challenge
Two identical systems: one for research, one for production Phase 1 (2013/2014): 2 x 475 Tflops peak▶ 2 x 990 bullx B710 DLC compute nodesPhase 2 (2016): 2 x 2.45 Pflops peak▶ 2 x 1800 bullx B720 DLC compute nodes
Solution
▶ More efficient forecasting processes (finer resolution, longer range)
▶ Improved forecasts for rain storms, wind in areas of high ground, low-level clouds and fog
▶ A bullx configuration that optimizes power consumption, footprint and cooling with patented Direct Liquid Cooling
Benefit
▶ Strong availability of the 2 supercomputers
>>
99,75%▶ Per month
>
3M jobs
Earlyintegration
7The migration
from DMF to HPSS (15PB)
was completed
after only 7 months
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Cellule de veille technologique
▶ Anticiper l’arrivée d’architectures Pré Exascale
– Préparer les communautés scientifiques françaises
– Fédérer une expertise HPC françaiseOpen and multi-technology
▶ Suivre et anticiper architectures HPC émergentes
▶ Mettre à disposition petits systèmes de test pour
évaluer architectures les + pertinentes
– Financement de petits matériels, logiciels et support applicatif
– Accès national à des système Ultra-energy efficient
▶ Mise en place d’une antenne du CEPP à Montpellier
Applications candidates• Climat : DYNAMICO et MesoNH• CFD : YALES2 et TRIO-CFD• Ingénierie : PATMOS• Sismique : SPECFEM3D• Astrophysique : RAMSES-GPU et hydro
• Physique haute énergies : GYSELA, SMILEI et deuxapplications IN2P3• Matériaux : Metawalls• Physique / Chimie : BigDFT et QMC=Chem• Maths : Sparse multifrontal QR et NT2
4ièmes Journées Scientifiques Equip@Meso : Sciences de l’Univers, Toulouse, CALMIP 26&27 Novembre 2015
Conclusion