Paving The Road to Exascale Computing · 2020. 1. 14. · •Topology aware collectives take advantage of optimized message coalescing •Makes use of powerful multicast capabilities

© 2012 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 1

Paving The Road to Exascale Computing

March 2012, [email protected]


Connectivity Solutions for Efficient Computing

Enterprise HPC

High-end HPC

Leading Connectivity Solution Provider For Servers and Storage

HPC Clouds

Host/Fabric

Software

Mellanox Interconnect Networking Solutions

ICs Switches/Gateways Adapter Cards Cables


Mellanox Complete End-to-End Connectivity

- Collectives Accelerations (FCA/CORE-Direct)

- GPU Accelerations (GPUDirect)

- MPI/SHMEM/PGAS

- RDMA

- Quality of Service

- Adaptive Routing

- Congestion Management

- Fat-Tree, 3D Torus

- UFM, Mellanox OS

- Integration with job schedulers

- Inbox Drivers

Server and Storage High-Speed Connectivity

Networking Efficiency/Scalability

Application Accelerations

Host/Fabric Software Management

- Latency

- Bandwidth

- CPU Utilization

- Message rate


Paving The Road to Exascale Computing

Mellanox InfiniBand is the interconnect of choice for PetaScale computing

• Accelerating 50% of the sustained PetaScale systems (5 systems out of 10)

Dawning (China) TSUBAME (Japan) NASA (USA)

>11K nodes

CEA (France) LANL (USA)


ORNL “Spider” System – Lustre File System

Oak Ridge Nation Lab central storage system • 13400 drives

• 192 Lustre OSS

• 240GB/s bandwidth

• Mellanox InfiniBand interconnect

• 10PB capacity

World leading high- performance InfiniBand storage system


Introducing FDR InfiniBand 56Gb/s Solutions

2002

10Gb/s

2005

20Gb/s

2008

40Gb/s

2011

56Gb/s

3.0

Highest Performance, Reliability, Scalability, Efficiency


Mellanox Virtual Protocol Interconnect Solutions with PCIe 3.0

Industry leader • PCI Express 3.0 • Dual-port FDR IB or 40GbE • Native RDMA • CORE-Direct

Industry leader

• 36 x FDR IB or 40GE 64 x 10GbE

• Integrated routers and bridges

• 4Tbit switching capacity

• Ultra-low latency

• Switch systems: from 36-port to 648-port

Networking Storage Clustering Management

Applications

Adapter Card

Mezzanine Card

Switch OS Layer

Unified Fabric Manager

SW Acceleration Products

LOM

http://www.redhat.com/


Reliability / Efficiency

• Link bit encoding – 64/66

• Forward Error Correction

• Lower power consumption

Performance / Scalability

• >12GB/s bandwidth, <0.7usec latency

• PCI Express 3.0

• InfiniBand router and IB-Eth bridges

FDR InfiniBand New Features and Capabilities


Modular Switch

Edge Switch

SX6500

Up to 648 ports FDR

FDR 56Gb/s InfiniBand Solutions Portfolio

Management

Feature License

Keys

Cables

Mellanox M-1

E-2-E Cluster

Support Services

UFM Diagnostics

Note: Please check availability with your Mellanox representative

Adapters

SX6025 – 36 ports externally managed

SX6036 – 36 ports managed

SX60XX

1U 36 port

Single and dual port

FDR adapter card

Virtual Protocol Interconnect

648p 324p 108p 216p


Double the Bandwidth, Half the Latency

120% Higher Application ROI

FDR InfiniBand PCIe 3.0 vs QDR InfiniBand PCIe 2.0


FDR InfiniBand with PCIe 3.0 In Multiple Deployments

FDR InfiniBand is The New Deal!

187Tflops with Only 648 Nodes

#54 on the TOP500 list

And

More!

http://rds.yahoo.com/_ylt=A2KJke6djfJNdnwA3fOjzbkF/SIG=12cmio2tj/EXP=1307770397/**http:/www.aei.mpg.de/~allen/TestBedWeb/PR/logos/LRZ.gif


Mellanox offers the most scalable and complete parallel programming solution • ScalableMPI

• ScalableSHMEM

• ScalableUPC

Mellanox Fabric Collective Accelerations (FCA) • Topology aware collectives take advantage of optimized message coalescing

• Makes use of powerful multicast capabilities in network for one-to-many communication

• Run collectives on separate service level so no interference with other communications

• Utilizes Mellanox CoreDirect collective hardware offload to minimize system noise

Mellanox Message Accelerations (MXM) • Delivers highest performing and scalable parallel programming interface in the industry

• Implements reliable messaging optimized for Mellanox HCAs

• Hybrid transport services

• Supports one-sided communications

• Intra-node shared memory

• Efficient memory registration and management

Mellanox ScalableHPC Programming Libraries


Mellanox MPI Accelerations


Mellanox MPI Scalability


Mellanox MPI optimization enable linear strong scaling for LLNL application

World Leading Performance and Scalability

Mellanox MPI Optimization – Highest Scalability at LLNL


Mellanox ScalableSHMEM Collective Scalability


Thank You [email protected]

Paving The Road to Exascale Computing · 2020. 1. 14. · •Topology aware collectives take advantage of optimized message coalescing •Makes use of powerful multicast capabilities

Documents