Architect of
an Open World™
Extreme Computing: The Bull Way
Dr.-Ing. Joachim Redmer, Director HPC
(
)
Bull is an Information Technology
company, focusing on open and
secure systems
Our mission is to help corporations
and public sector organizations
optimize the architecture,
operations and financial return of
their Information Systems,
supporting their core business
processes
Bull is the only European IT
company that is positioned to
deliver all the key elements of the
IT value chain
2009 Bull activities
w/o consolidation of Amesys activities
3 ©Bull, 2011 Bull Extreme Computing
Reached 100 M€ products revenue in 2009
-
On track to exceed 10% market share in Europe
Anticipated 150 M€ products revenue in 2010
Introduced bullx range in 2009
-
Delivering new levels of performance and innovation for
Extreme Computing
-
Xeon-based as well as hybrid blades (NVIDIA)
-
bullx nominated best HPC server product by HPCWire
Signed landmark deals for bullx servers in 2010
-
GENCI /TGCC (France)
-
AWE - United Kingdom
-
RWTH Aachen - Germany
-
Dassault Aviation (France)
-
Société Générale (France)
-
Ineris (France)
-
Reims University (France), …
first petascale system in Europe at CEA
500 specialists dedicated to HPC in Europe
SOLUTIONS &
INTEGRATION
2008: SIRUS - Vertical ISV &
SI in France
2008: CSB Consulting - IT
consulting in Belux
2007: Siconet - SI in Spain
2006: Address vision - Postal
automation ISV & SI in USA
2006: AMG - Telco SI in
Poland
EXTREME COMPUTING
-
2008: Science +
Computing - Extreme
Computing Solutions &
Services in Germany
-
2007: Serviware - Extreme
Computing SI in France
SECURITY &
OUTSOURCING
-
2010: Amesys
-
2006: Agarik - Internet SI
& hoster in France
-
2005: Enatel
enatel
6 ©Bull, 2011 Bull Extreme Computing
Structural Mechanics
Implicit
Structural Mechanics
Explicit
Computational Fluid
Dynamics
Electro-Magnetics
Computational Chemistry
Quantum Mechanics
Reservoir Simulation
Rendering / Ray Tracing
Climate / Weather
Ocean Simulation
Data Analytics
Computational Chemistry
Molecular Dynamics
Computational Biology
Seismic Processing
eXtreme Computing applications
A dedicated team of experts in
application performance
Architect of
an Open World™
Architect of
an Open World™
2010: TERA 100
1.25
Peak PFlops
1.05
Linpack PFlops
4 300
bullx S nodes
140 000
Intel Nehalem-EX cores
300
TB of memory
20
PB of disk storage
QDR InfiniBand interconnect
500
GB/s bandwidth to the global file system
Bullx Supercomputer: Best Top10 Linpack efficiency
0,3
0,4
0,5
0,6
0,7
0,8
0,9
10 ©Bull, 2011 Bull Extreme Computing
"With technical support from CEA, through a
competitive tendering process, we were
able to assess the excellence of Bull's
offering. This means we will soon have at
our disposal a machine that will offer
French and European scientists the
resources they need to carry out their
research work at the highest possible
level in a highly competitive global
environment,"
Catherine Rivière, CEO of GENCI
1
st
phase Implemented in October
2010
CURIE in figures
1.6
PetaFlops
(90,000+ Xeon cores)
10
PB of storage
250
GB/s data throughput
200
m² footprint
Architect of
an Open World™
2011: RWTH Aachen
292
Peak TFlops
1350
bullx blades
362
bullx S nodes
16 200
Intel Westmere cores
11 500
Intel Nehalem-EX cores
1500
TB HPC storage
1500
TB Home storage
QDR InfiniBand interconnect
Architect of
an Open World™
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
15 ©Bull, 2011 Bull Extreme Computing
bullx blade system – overall concept
General purpose, versatile
-
Xeon Westmere processor
-
96 GB RAM per blade
-
Local HDD/SSD or Diskless
-
IB / GBE
-
RH, Suse, Win HPC2008, CentOs, …
-
Compilers: GNU, Intel, …
Uncompromised performances
-
Support of high frequency Westmere
-
Memory bandwidth: 12x mem slots
-
Fully non blocking IB QDR interconnect
-
2.64TFLOPS per chassis (Intel Xeon
X5675 3.06GHz)
-
Up to 15.8TFLOPS per rack (with CPUs)
High density
-
7U chassis
-
18x blades with 2 proc, 12x DIMMs,
HDD/SSD slot/IB connection
-
1x IB switch (36 ports)
-
1x GBE switch (24p)
-
10 GigE Uplink (optional)
-
Ultracapacitor
Leading edge technologies
-
Intel Westmere
-
InfiniBand QDR
-
Diskless
-
Ready for GPU blades
Optimized Power consumption
-
Typical 6.5 kW / chassis
-
High efficiency (90%) PSU
-
Smart fan control in each chassis
-
Smart fan control in water-cooled rack
17 ©Bull, 2011 Bull Extreme Computing
bullx blade system – Block Diagram
18x compute blades
-
2x Westmere-EP sockets
-
12x memory DDR3 DIMMs (12x 8GB= 96GB)
-
1x SATA HDD/SSD slot (optional – diskless
an option)
-
1x IB ConnectX/QDR chip
1x InfiniBand Switch Module (ISM) for
cluster interconnect
-
36 ports QDR IB switch
-
18x internal connections
-
18x external connections
1x Chassis Management Module (CMM)
-
OPMA board
-
24 ports GbE switch
-
18x internal ports to Blades
-
3x external ports
1x optional Ethernet Switch Module
(ESM)
-
24ports GbE switch
-
18x internal ports to Blades
-
3x external ports
1x optional Ethernet Switch Module
(TSM)
-
GigE Switch with 10GigE Uplinks
1x optional Ultra Capacitor Module
(UCM)
SATA
SSD
diskless
I/O Controller
(Tylersburg)
QPI
QPI
QPI
Westmere EP
Nehalem EP
Westmere EP
31.2GB/s
12.8GB/s
Each direction
31.2GB/s
QPI
QPI
PCIe 8x
4GB/s
QPI
Westmere EP
Nehalem EP
Westmere EP
31.2GB/s
12.8GB/s
Each direction
31.2GB/s
In
finiB
a
n
d
PCIe 16x
8GB/s
A
c
c
e
ler
a
to
r
PCIe 8x
4GB/s
In
finiB
a
n
d
PCIe 16x
8GB/s
A
c
c
e
le
ra
to
r
bullx B500 compute blade
bullx B505 accelerator blade
PCIe 8x
4GB/s
Inf
iniB
an
d
QPI
I/O
Controller
(Tylersburg)
I/O
Controller
(Tylersburg)
bullx blade system – blade block diagrams
SATA
SSD
diskless
20 ©Bull, 2011 Bull Extreme Computing
Ultracapacitor Module (UCM)
NESSCAP Capacitors (2x6)
Leds
Board
Embedded protection against short
power outages
Protect one chassis with all its
equipment under load
Up to 250ms
Avoid on site UPS
save on infrastructure costs
save up to 15% on electrical costs
Improve overall availability
bullx chassis packaging
LCD
unit
18x blades
7
U c
h
a
s
s
is
PSU x4
ESM/TSM
CMM
22 ©Bull, 2011 Bull Extreme Computing
bullx B505 accelerator blade
• 2 x Intel Xeon 5600
• 2 x NVIDIA T20
• 2 x IB QDR
7U
18.5 TFLOPS in 7 U
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
25 ©Bull, 2011 Bull Extreme Computing
R424 E2
bullx rack-mounted systems
Supports latest graphics
& accelerator cards
4U or tower
2-Socket
Xeon 5600
18 DIMMs
2 PCI-Express x16 Gen2
8x SATA2 or 8x SAS HDD
Powerful power supply
Hot-swap Fans
Enhanced connectivity
and storage
2U
Xeon 5600
2-Socket
18 DIMMs
2 PCI-Express x16 Gen2
8x SATA2 or 8x SAS
HDD
Redundant power supply
Hot-swap fans
4 nodes per 2U
for unprecedented density
Xeon 5600
2x 2-Socket
2x 12 DIMMs
QPI up to 6.4 GT/s
2x 1 PCI-Express x16 Gen2
InfiniBand QDR embedded
(optional)
3x SATA2 hot-swap HDD
92% PSU efficiency
CO
M
PUTE
NO
DE
R423 E2
SER
VICE
NO
DE
R425 E2
VISUA
LI
ZA
TI
O
N
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
27 ©Bull, 2011 Bull Extreme Computing
Mesca’s fundamentals
SMP of up to 16 sockets based on Bull Coherent Switch
-
Intel Xeon Nehalem EX processors
-
Coherent memory of up to 1 TB
Several types of packaging
-
High-density compute node
-
High I/O connectivity node
RAS features
-
Self-healing of the QPI and XQPI
-
Hot swap disk, fans, power supplies
Green features
Module level diagrams
Full width QPI link
IOH
IOH
For 4 socket-only systems
NHM
NHM
NHM
NHM
XCSI link
Repeated n times
for >4 socket systems
NHM
NHM
NHM
NHM
IOH
IOH
BCS
29 ©Bull, 2011 Bull Extreme Computing
Bullx S60x0 – CC-NUMA server
Node maximum configuration :
4 modules
16 sockets
128 cores (Nehalem-EX)
128 memory slots (2TB)
BCS
BCS
BCS
BCS
NHM
EX
NHM
EX
NHM
EX
NHM
EX
IOH
IOH
BCS
SMP (CC-NUMA)
Large nodes :
Large shared memory (pre/post-processing)
Many more cores (SMP)
Fewer nodes
Simpler system administration
Taking advantage of SMP nodes with MPI
Optimized intra-node
throughput
-
Enable direct copy from sender
to receiver
-
Rely on SSE instruction set
-
Achieve a transfer rate of half
the memory bandwidth
Optimized intra-node latency
-
Lock-free shared memory device
-
Take advantage of socket
architecture
-
Shared cache latency: 200 nsec
00,2 0,4 0,6 0,8 1 1,2 1,4 1,6 1,8 4 cores, Mvapich2 8 cores, Mvapich2 16 cores, Mvapich2 4 cores, MPIBull2 8 cores, MPIBull2 16 cores, MPIBull2 To socket 1, core 2 To socket 1, core 3 and 4 To socket 2
31 ©Bull, 2011 Bull Extreme Computing
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
Bull Storage for HPC clusters
*: with Lustre
A complete line of
storage systems
• Performance
• Modularity
• High Availability*
A rich
management suite
• Monitoring
• Grid & standalone
system deployment
• Performance
33 ©Bull, 2011 Bull Extreme Computing
Bull Storage Systems for HPC
StoreWay
Optima 1500
DataDirect Networks
SFA 10k
(consult us)
SAS/SATA
3 to 144 HDDs
Up to 12 host ports
2U drawers
SAS/SATA
Up to 1200
HDDs
8 host ports
4+ 2/3/4 U
drawers
FC/SATA
Up to 480 HDDs
Up to 16 host ports
3U drawers
StoreWay
EMC CX4
cluster
suite
Bull Storage Systems for HPC - details
Optima1500
CX4-120
CX4-480
SFA 10k
couplet
#disk
144
120
480
1200
Disk type
SAS 146/300/450 GBSATA 1TB FC 146/300/400/450 GB SATA 1TB FC 10Krpm 400 GB 15Krpm 146/300/450 GB SATA 1TB SAS 10Krpm 400 GB SAS 15Krpm 300/450/600 GB SATA 1000/2000 GB
RAID
R1, 3, 3DP, 5, 6, 10, 50 and
TM
R0, R1, R10, R3, R5, R6
R0, R1, R10, R3, R5, R6
8+2 (RAID 6)
Host ports
2/12 FC 4
4/12 FC4
8/16 FC4
16 FC8 / 8 QDR
Back end ports
2 SAS 4X
2
8
20 SAS 4X
Cache size
(max)
4 GB
6GB
16GB
5 GB
RAID-protected
Controller size
2 U base with disks
3 U
3 U
4 U
Disk drawer
2 U
12 slots
3 U
15 slots
3 U
15 slots
4 U
60 slots
Performance
(
MB/s; Raid5)
R: Read; W:Write
R: up to 900 MB/s W: up to 440 MB/s R: up to 720 MB/s W: up to 410 MB/s R: up to 1.25 GB/s W: up to 800 MB/s R&W: up to 20 GB/s35 ©Bull, 2011 Bull Extreme Computing
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
Bull Cool Cabinet door
No impact on server behaviour
air flow through doors is adjusted to match
the drawer air flows
No impact on computer room
Select outlet Air temperature
2 way or 3 way valve controls heat
exchanger water flow.
No more hot spots
better MTBF
Cools up to 40kW per rack
ready for Bull Extreme Computing systems
Inlet water 7°C
Server
ΔT 15°CServer
ΔT 15°CServer
ΔT 15°CServer
ΔT 15°CServer
ΔT 15°C Room pressureWater cooled door
M Δ p Rack pressure 35°C 20°C E x ch a n g e r 20°C
37 ©Bull, 2011 Bull Extreme Computing
Cool cabinet door: Characteristics
Width
600mm (19
”
)
Height
2020mm (42U)
Depth
200mm (8
”
)
Weight
150 kg
Cooling capacity
Up to 40 kW
Power supply
Redundant
Power consumption
700 W
Input water temperature
7-12
°
C
Output water temperature
12-17
°
C
Water flow
2 liter/second (7 m3/hour)
Ventilation
14 managed multi-speed fans
Recommended
cabinet air inlet
20
°
C +- 2
°
C
Cabinet air outlet
20
°
C +- 2
°
C
Management
Integrated management board for local regulation
39 ©Bull, 2011 Bull Extreme Computing
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
bullx supercomputer suite
Standard
Edition
Advanced Edition
Advanced Edition
eXtreme Pack
thousands of nodes
41 ©Bull, 2011 Bull Extreme Computing
42 ©Bull, 2011 Bull Extreme Computing
Advanced Edition / Extreme Pack: components
• Super-Fast image based provisioning
• Web-based Multi-level supervision
• Power management
• Automated health management
• Maintenance management
bullx MC
• Highly available cells based architecture
• Increased throughput and scalability
bullx PFS
• Advanced placement policies
• Topology aware resource allocation
bullx BM
• Multi-path network failover
• Abnormal patterns detection
• Topology aware operations
bullx MPI
• Complete best of breed set of tools (from compiling,
debugging to profiling and optimizing activities)
bullx DE
• HPC Enabled (OS jitter reduction, Optimized operations for
increased application performance)
• Enhanced OFED
bullx Linux
Management Center
Parallel File System
Batch Management
44 ©Bull, 2011 Bull Extreme Computing
bullx blade system
bullx rack-mounted systems
bullx SMP system
NVIDIA Tesla Systems
Bull Storage
Cool cabinet door
bullx cluster suite
Windows HPC Server 2008
Bull and Windows HPC Server 2008
Clusters of bullx R422 E2 servers
-
Intel® 5600 processors
-
Compact rack design: 2 servers in 1U
-
Fast & reliable InfiniBand interconnect
supporting
Microsoft
®
Windows HPC Server 2008
-
Simplified cluster deployment and management
-
Broad application support
-
Enterprise-class performance and scalability
Common collaboration with leading ISVs to provide
complete solutions
The right technologies to handle
industrial applications efficiently
46 ©Bull, 2011 Bull Extreme Computing
Windows HPC Server 2008
Microsoft® Windows
Server® 2008 HPC
Edition
Microsoft® HPC
Pack 2008
+
=
Microsoft®
Windows® HPC
Server 2008
•
Support for high
performance
hardware (x64 bit
architecture)
•
Winsock Direct
support for
RDMA for high
performance
interconnects
(Gigabit Ethernet,
InfiniBand, Myrinet,
and others)
•
Support for Industry
Standards MPI2
•
Integrated Job
Scheduler
•
Cluster Resource
Management Tools
•
Integrated “out of the
box” solution
•
Leverages past
investments in
Windows skills and
tools
•
Makes cluster
operation just as
simple and secure as
operating a single
system
Combining the power of the Windows Server platform
with rich, out-of-the-box functionality to help improve the productivity
and reduce the complexity of your HPC environment
A complete turn-key solution
Bull delivers a complete ready-to-run solution
-
Sizing
-
Factory pre-installed and pre-configured
(R@ck’n Roll)
-
Installation, integration in the existing infrastructure
-
1st and 2nd level support
-
Monitoring, audit
-
Training
48 ©Bull, 2011 Bull Extreme Computing
bullx cluster 400-W
bullx cluster 400-W4
-
4 compute nodes to relieve the
strain on your work stations
bullx cluster 400-W8
-
8 compute nodes to give
independent compute resources to
a small team of users, enabling
them to submit large jobs or
several jobs simultaneously
bullx cluster 400-W16
-
16 compute nodes to equip a
workgroup with independent high
performance computing resources
that can handle their global
compute workload
A solution that combines:
The performance of bullx rack
servers equipped with Intel
®
Xeon
®
processors
The advantages of Windows HPC
Server 2008
-
Simplified cluster deployment and
management
-
Easy integration with IT
infrastructure
-
Broad application support
-
Familiar development environment
And expert support from Bull
’
s
Microsoft Competence Center
Enter the world of High Performance Computing with