Huawei Rozwiązania i platformy HPC

Transcription

Huawei Rozwiązania i platformy HPC
Huawei
Rozwiązania i platformy
HPC
Dominik Dziarczykowski
HUAWEI TECHNOLOGIES CO., LTD.
www.huawei.com
Huawei IT Portfolio
Converged
Infrastructure
Worldwide
FusionInsight
E9000
X8000
Rack Server
FusionSphere
OceanStor
18000 Series
X6800
FusionAccess
OceanStor
9000
OceanStor
N8500
OceanStor
Dorado5100
X6000
Enterprise
4U 4S
FusionCube
RH8100
2U 2S
OceanStor
Dorado2100 G2
RH5885
1U 2S
RH2288
RH1288
SME/Branch
ES3000
SSD Card
OceanSto
rS2200T
ManageOne
DC Management
Cloud
Datacenter
OceanStor
OceanStor 5800 V3
5600 V3
OceanStor
5300/5500 V3
OceanSto
r S2600T
Container DC
Modular DC
Micro DC
Scalability/Reliability
Storage HPC Solutions
Typical Scenario
Big Capacity Scenario
High I/O Scenario
High Bandwidth Scenario

Integration Bandwidth < 20GB/s

Integration Bandwidth < 100GB/s

Integration Bandwidth <200GB/s

Integration Bandwidth <500GB/s

Single node < 2 GB/s

Single node < 4GB/s

Single node < 800MB/s

Single node 3,6,9-12GB/s

About 100K files

About 100K files

More than 500K files

More than 500K files

For big capacity


For balanced performance and
capacity


MDS
OSS
OSS
Lustre
parallel file system
High I/O for creating lots of files
per second
For big capacity and high
performance
For high performance
OSS
MDS
OSS
IB/10GE
IB/10GE
Lustre
parallel file system
IB
SAN
…
…
…
…
…..
OceanStor V3
FC
FusionStorage
IP/IB
IP/IB
OceanStor 9000 Wushan FS
Xyratex ClusterStor
IP/IB
IP/IB
Huawei HPC Solution: Network and Infrastructure Options
Ethernet Switches
IB Switches
Equipment Room
Switches
CE12812
CE12808
CE12804
CE6800
CE5800
Mini (< 20 m2)
Application scenario: small-sized
enterprises
Deployment scale: 1-5 server cabinets
Small (30-100 m2)
Application scenario: small- and mediumsized enterprises
Deployment scale: 6-28 server cabinets
Mellanox :MIS51
Mellanox:
xx MIS52xx
MIS50xx MIS60xx
MIS65xx
Intel 12200
……
Medium and large (100-2000+ m2)
Application scenario: medium- and largesized enterprises
Deployment scale: 28-1000+ server
Elastic HPC Solution
Cloud Services
Expanding flexibly
Amazon
Customizing Portal (SI or Customer)
BCM providing unified scheduling
VM (Virtual machines)
BCM monitoring and
supervising
PM (Physical machines)
FusionSphere / Openstack
Storage
OceanStor 9000
Dorado5100
OceanStor V3
Data center Network
Network & Security
CE Series
Infiniband
ManageOne
HPC Hardware Pool
Server
E9000
RH5885/H
RH2288
HUAWEI HPC Sucessful Cases
Central and Eastern Asia

Europe
Turkish Academic Network and Information
Centre
North America

Yildiz Technical University

Istanbul Technical University in Turkey

Harran University

State University of Iowa

University of Nebraska-Lincoln

University of Tennessee, Knoxville

University of Santa Cruz

Digital domain

Deltares Institute in the Netherlands

Sankt-Petersburg university in Russia

Italian CNR

Utility Association Eilenburg-Wurzen in

PCSS in Poland

ICM University of Warsaw

Illumination in France
German

Japan

Newcastle University in U.K. (Phases
I,II,III,IIII)
Kyushu University
China

Institute of Disaster Prevention

institution
Science and Technology

Latin America
Asia

Mexico Water Conservancy Bureau

Mexico Ministry of agriculture

Mackenzie University in Brazil

Astronomy Institute in Chile

Institute of Medical foundation of University of Sao

Provincial environment
China electricity research

Shanghai Life sciences research
protection bureau in HeNan
institution of China science
Beijing Data Research
institution

Meteorological bureau in Macao
Department, Post and

Sino-Singapore Tianjin Eco-city

Singapore Global Foundries
Telecommunications Scientific

Tsinghua university

Meteorological bureau in Philippines
Institut

Beijing Jiaotong University

Beihang University

School of Psychology, Southwest

Capital Medical University

Changsha national
Paulo (USP) in Brazil
University

institution
supercomputing center
Case in 2013
Case in 2014

China electricity research
Guangzhou national

Beijing Forestry University
supercomputing center

Provincial environment protection
bureau in HeBei
PCSS-Poznanskie Centrum Superkomputerowo-Sieciowe
.
“PCSS” in English “PSNC”(Poznan Supercomputing and Networking Center).
Member of PL Grid organisation, leaders in High performance computing field and energy
efficient solutions.
PCSS provides service to all scientist in Poland and worldwide.
Solution Scope
9 x E9000 Blade Chassis
138 CH121V3 Blades
server with Intel Haswell
E5-2697 v3 CPU
Based on Huawei E9000
Blade server, partner
customized Direct liquid
cooling system with water,
improve cooling efficiency
Results
Average efficiency of
node 0,901Tflops
Hardware with ~60 kW
power consumption
80% of the heat removed by
the water loop
HPL test: only 2 servers
failed or were DOA
ICM-Uniwersytet Warszawski
.
Solution Scope
ICM, Interdisciplinary Centre for Mathematical and Computational Modeling, Warsaw
University.
Leading scientific high-end computing centre in Poland. ICM is member of Poland PL-Grid
Group in High performance computing field, provide the service for scientists.
Why Huawei
15 x E9000 Blade
Chassis
High Reliability
240 CH121V3 Blades
server with Intel
Haswell E5-2697 v3
CPU
Energy efficiency and
telecom class quality
Cluster reach a
theoretically
279.5TFLOPS
computing capability
Huawei provides overall
low TCO solution
https://www.youtube.com/watch?v=cUM4hDC-3dI
European Organization for Nuclear Research (CERN)
Explores the Particles with OceanStor
Customer Challenges
•
•
•
Over 25 PB of annual data growth
High reliability and cross-region data cloud sharing
Long-term data storage and maximum total cost of ownership
Huawei Solution
• Applies UDS' EB-level storage capacity to meet the growing data
storage requirements.
• Applies multiple data copies to enhance reliability of service data.
"CERN is hitting the technology limits for
resource-intensive simulations and analysis. Our
collaboration with Huawei shows an exciting new
approach, where their novel architecture extends
the capabilities in preparation for the exascale data rates
and volumes we expect in the future."
----- Bob Jones, Head of CERN openlab.
Customer Benefits
•
•
•
TCO reduced by 45%
High reliability with no data loss
Distributed architecture and EB-level storage capacity meeting the
storage requirements for the following 40 years
UCSC Explores the Mysteries of Heavenly Bodies with HUAWEI UDS
Challenges

The capacity of the existing storage system failed to meet the requirements of
astronomical observation data and modeling data that were growing at a rate
of PB.

The traditional storage system caused expensive purchase and O&M costs.

An isolated data storage platform was incapable of sharing results among
colleges.
Huawei Solution
“Hyades is more than ten times better than our
previous machine, and with the Huawei system
providing storage for our simulation results, we
can maximize the value of those results by
making them available to the astrophysics
community.”
Piero Madau, professor of astronomy and astrophysics
and principal investigator on the National Science
Foundation (NSF) grant

The UDS' EB-level capacity met the growing data storage requirements.

A high-density and low-consumption hardware platform and ZeroTouch
reached a balance between data stability and storage utilization.

Open S3 APIs and a unified namespace enabled multi-point access and
resource sharing
Customer Benefits

The distributed architecture and EB-level storage capacity made it easy to
handle data explosion.

An efficient, low-consumption, and intelligent design lowered 45% of TCO.

A unified resource pool and a cross-region data sharing platform improved
data access efficiency.

The on-demand dynamic storage space allocation adapted to growing services.