paving the road to exascale computing - hpc advisory council

15
Paving the Road to Exascale Computing Yossi Avni [email protected]

Upload: others

Post on 24-Mar-2022

3 views

Category:

Documents


0 download

TRANSCRIPT

Paving the Road to Exascale

Computing

Yossi Avni

[email protected]

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 2

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

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 3

Complete End-to-End Connectivity

- Collectives Accelerations (FCA/CORE-Direct)

- GPU Accelerations (GPUDirect)

- MPI/SHMEM

- RDMA

- Quality of Service

- Adaptive Routing

- Congestion Management

- Traffic aware Routing (TARA)

- UFM, FabricIT

- 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

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 4

Mellanox’s Interconnect Leadership

Highest

Performance

Complete

Eco-System

End-to-End

Quality

Advanced

HPC

MPI/SHMEM Offloads

GPU Acceleration

Adaptive Routing

Highest Throughput

Lowest Latency

CPU Availability

Auto Negotiation

Cable Reach

Signal Integrity

Power Management

Congestion Control

Message Rate

Topologies/Routing

From Silicon to

System

RDMA

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 5

2005 - 2006 - 2007 - 2008 - 2009 - 2010 - 2011

Ban

dw

idth

p

er

dir

ecti

on

(G

b/s

)

40G-IB-DDR

60G-IB-DDR

120G-IB-QDR

80G-IB-QDR

200G-IB-EDR

112G-IB-FDR

300G-IB-EDR

168G-IB-FDR

8x HDR

12x HDR

# of Lanes per direction

Per Lane & Rounded Per Link Bandwidth (Gb/s)

5G-IB DDR

10G-IB QDR

14G-IB-FDR (14.025)

26G-IB-EDR (25.78125)

12 60+60 120+120 168+168 300+300

8 40+40 80+80 112+112 200+200

4 20+20 40+40 56+56 100+100

1 5+5 10+10 14+14 25+25

x12

x8

x1

2014

12x NDR

8x NDR

40G-IB-QDR

100G-IB-EDR

56G-IB-FDR

20G-IB-DDR

4x HDR

x4

4x NDR

10G-IB-QDR

25G-IB-EDR

14G-IB-FDR

1x HDR

1x NDR

Market

Demand

InfiniBand Link Speed Roadmap

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 6

Next Generation InfiniBand Technology

Highest Throughput Connectivity for Server and Storage

Available: 2011 (end-to-end: adapters, switches, cables)

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 7

Offloading at the HCA

(CORE-Direct)

Offloading at the

network/switches

(iCPU)

Node

Scalable MPI Collectives Acceleration with FCA

Offloading/acceleration

management (FCA)

~20% Performance increase at 16 nodes!

Most Scalable Offloading for MPI Applications

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 8

Mellanox Message Rate Performance Results

PPN – process per node, or cores per node

Highest MPI Message Rate!

90 Million messages per Second

Highest IB Message Rate!

23 Million messages per Second

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 9

Network Utilization via Traffic Aware Routing

Maximizing Network Utilization

Job Submitted in Scheduler Matching Jobs Automatically

Application Level Monitoring

& Optimization Measurements

Fabric-wide Policy Pushed to Match

Application Requirements

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 10

Hardware Based Congestion Control

Network Latency % Improvement

Ping Pong Latency 88%

Natural Ring Latency 81.6%

Random Ring Latency 81.3%

Ping Pong bandwidth 85.5%

Applications (HPCC) % Improvement

PTRANS 76%

FFT 40%

For more performance examples:

“First Experiences with Congestion

Control in InfiniBand Hardware”;

Ernst Gunnar Gran, Magne Eimot,

Sven-Arne Reinemo, Tor Skeie,

Olav Lysne, Lars Paul Huse, Gilad

Shainer; IPDPS 2010

Congestion Free Network For Highest Efficiency

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 11

Highest Performance GPU Clusters with GPUDirect

GPUDirect

GPU computing mandates Mellanox solutions

GPUDirect: 35% application performance increase

3 nodes 3 nodes

Mellanox InfiniBand Accelerates GPU Communications

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 12

Superior InfiniBand Solutions

Performance: 45% Lower latency, highest throughput and 3x the message rate

Scalability: proven for Petascale computing, highest scalability through accelerations

Reliability :from silicon to system, highest signal integrity, two order of magnitude lower BER

Efficiency: Highest CPU/GPU availability through complete offloading, low power consumption

Certification: Complete ISVs support and qualification, MPI vendors, job schedulers

Return on investment: Most cost/effective, simple to manage, 40Gb/s end-to-end connectivity

Financial Mellanox Connectivity Solutions

Financial

Cloud &

Web 2.0 Clustered

Database

Weather

Digital Media

Computational Aided Engineering

Bioscience

Labs, Research

University, Academic

Oil and Gas

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 13

Complete High-Performance Scalable Interconnect

Solutions for Server and Storage

Bottom Line Mellanox Benefits for HPC

TCO

50+% Reduction

Energy Costs

65+% Reduction

Infrastructure

60+% Saving

Performance

100+% Increase

Entry-level HPC High-end HPC Enterprise HPC HPC Clouds

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 14

9 of Top 10 Automotive

Manufacturers

4 of Top 10 Pharmaceutical

Companies

7 of Top 10 Oil and Gas Companies

6 of Top 10 Global Banks

30%+ of Fortune-100 & top global-High Performance Computers

Performance Leadership Across Industries

© 2011 MELLANOX TECHNOLOGIES - MELLANOX CONFIDENTIAL - 15

Thank You

Thinking, Designing and Building Scalable HPC

[email protected]