Contenu connexe Similaire à Mellanox hpc day 2011 kiev Similaire à Mellanox hpc day 2011 kiev (20) Plus de Volodymyr Saviak (15) Mellanox hpc day 2011 kiev1. Paving The Road to Exascale Computing
The Highest Performance, Most Scalable
Interconnect Solutions for Servers and Storage
Oct 2011, HPC@mellanox.com
2. Connectivity Solutions for Efficient Computing
Enterprise HPC
High-end HPC HPC Clouds
Mellanox Interconnect Networking Solutions
Host/Fabric
ICs Adapter Cards Switches/Gateways Cables
Software
Leading Connectivity Solution Provider For Servers and Storage
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 2
3. Mellanox Complete End-to-End Connectivity
Host/Fabric Software Management
- UFM, Mellanox OS
- Integration with job schedulers
- Inbox Drivers
- Collectives Accelerations (FCA/CORE-Direct)
Application Accelerations - GPU Accelerations (GPUDirect)
- MPI/SHMEM/PGAS
- RDMA
- Quality of Service
- Adaptive Routing
Networking Efficiency/Scalability - Congestion Management
- Traffic aware Routing (TARA)
Server and Storage High-Speed Connectivity
- Latency - CPU Utilization
- Bandwidth - Message rate
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 3
4. Interconnect Trends – Top100, Top200, Top300, Top400
InfiniBand connects the majority of the TOP100, 200, 300, 400 supercomputers
• 42% of the Top500 systems
Due to superior performance, scalability, efficiency and return on investment
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 4
5. Powering the Petascale Today
Dawning (China) TSUBAME (Japan) NASA (USA)
>11K nodes
LANL (USA) CEA (France)
InfiniBand is the interconnect of choice for PetaScale computing, all Mellanox
• Accelerating 50% of the sustained PetaScale systems (5 systems out of 10)
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 5
7. Introducing FDR InfiniBand 56Gb/s Solutions
2011
2008 56Gb/s
2005 40Gb/s
2002
20Gb/s
10Gb/s
3.0
Highest Performance, Reliability, Scalability, Efficiency
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 7
8. Recent Introductions
Unified Fabric Manager
Applications Switch OS Layer
Networking Storage Clustering Management
SW Acceleration Products
LOM
Adapter Card
Industry leader Mezzanine Card Industry leader
• PCI-express gen3 • 36 x FDR IB or 40GE 64 x 10GE
• Dual-port FDR IB or 40GE • Integrated routers and bridges
• Native RDMA • 4Tbit switching capacity
• Core-Direct • Ultra-low latency
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 8
9. FDR InfiniBand New Features and Capabilities
Performance / Scalability Reliability / Efficiency
• Adapter / Switch port bandwidth 56Gb/s • Link bit encoding – 64/66
• Latency reduction • Forward Error Correction
• InfiniBand router and IB-Eth/FC bridges • Lower power consumption
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 9
11. Mellanox Collectives Acceleration Components
CORE-Direct
• Adapter-based hardware offloading for collectives operations
• Includes floating-point capability on the adapter for data reductions
• CORE-Direct API is exposed through the Mellanox drivers
- Available for 3rd party libraries/software protocols
FCA
• FCA is a software plug-in package that integrates into available MPIs
• FCA replaces the MPI software library code for collective communications
• FCA implements MPI collectives operations using the hardware accelerations
- Utilizing CORE-Direct (adapter offloading)
• FCA includes support for sophisticated collectives algorithms
• FCA is available through licensing
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 11
12. FCA/CORE-Direct – Application Performance
*Acknowledgment: HPC Advisory Council for providing the performance results
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 12
13. GPUDirect for GPU Accelerations
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 13
14. GPUDirect – Efficient GPU/Network Interface
Receive Transmit
1 2
2 1
System System
CPU CPU
Memory Memory
Chip Chip
GPU set GPU
set
InfiniBand InfiniBand
GPU
GPU Memory
Memory
GPUDirect
System System
1
1
Memory CPU CPU Memory
Chip Chip
GPU set set GPU
InfiniBand InfiniBand
GPU GPU
Memory Memory
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 14
15. GPUDirect – Application Performance
LAMPS
• 3 nodes, 10% gain
3 nodes, 1 GPU per node 3 nodes, 3 GPUs per node
Amber – Cellulose
• 8 nodes, 32% gain
Amber – FactorX
• 8 nodes, 27% gain
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 15
16. Thank You
HPC@mellanox.com
© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 16