Interconnect Your Future HPC... · X86, Power, GPU, ARM and FPGA-based Compute and Storage...

Paving the Road to Exascale

February 2016

Interconnect Your Future

The Ever Growing Demand for Higher Performance

2000 202020102005

“Roadrunner”

Terascale Petascale Exascale

Single-Core to Many-CoreSMP to Clusters

Performance Development

Co-Design

Hardware

Software

Application

The Interconnect is the Enabling Technology

Co-Design Architecture to Enable Exascale Performance

CPU-Centric Co-Design

Limited to Main CPU Usage

Results in Performance Limitation

Creating Synergies

Enables Higher Performance and Scale

Software

In-CPU

ComputingIn-Network

Computing

In-Storage

Computing

The Intelligence is Moving to the Interconnect

Interconnect

Past Future

Intelligent Interconnect Delivers Higher Datacenter ROI

Intelligence

Network Offloads

Computing for applications

Smart Network

Increase Datacenter Value

Network functions

On CPU

Breaking the Application Latency Wall

Today: Network device latencies are on the order of 100 nanoseconds

Challenge: Enabling the next order of magnitude improvement in application performance

Solution: Creating synergies between software and hardware – intelligent interconnect

Intelligent Interconnect Paves the Road to Exascale Performance

10 years ago

microsecond

NetworkCommunication

Framework

microsecond

Communication

Framework

microsecond

Network

microsecond

Communication

Framework

Future

microsecond

Co-Design

Network

Introducing Switch-IB 2 World’s First Smart Switch

The world fastest switch with <90 nanosecond latency

36-ports, 100Gb/s per port, 7.2Tb/s throughput, 7.02 Billion messages/sec

Adaptive Routing, Congestion control, support for multiple topologies

World’s First Smart Switch

Build for Scalable Compute and Storage Infrastructures

10X Higher Performance with The New Switch SHArP Technology

SHArP (Scalable Hierarchical Aggregation Protocol) Technology

Delivering 10X Performance Improvement

for MPI and SHMEM/PAGS Communications

Switch-IB 2 Enables the Switch Network to

Operate as a Co-Processor

SHArP Enables Switch-IB 2 to Manage and

Execute MPI Operations in the Network

SHArP Performance Advantage

MiniFE is a Finite Element mini-application

• Implements kernels that represent

implicit finite-element applications

10X to 25X Performance Improvement

AllRedcue MPI Collective

Multi-Host Socket DirectTM – Low Latency Socket Communication

Each CPU with direct network access

QPI avoidance for I/O – improve performance

Enables GPU / peer direct on both sockets

Solution is transparent to software

CPU CPUCPU CPUQPI

Multi-Host Socket Direct Performance

50% Lower CPU Utilization

20% lower Latency

Multi Host Evaluation Kit

Lower Application Latency, Free-up CPU

Introducing ConnectX-4 Lx Programmable Adapter

Scalable, Efficient, High-Performance and Flexible Solution

Security

Cloud/Virtualization

Storage

High Performance Computing

Precision Time Synchronization

Networking + FPGA

Mellanox Acceleration Engines

and FGPA Programmability

On One Adapter

Mellanox InfiniBand Proven and Most Scalable HPC Interconnect

“Summit” System “Sierra” System

Paving the Road to Exascale

20K InfiniBand nodes

Mellanox end-to-end scalable EDR, FDR and QDR InfiniBand

Supports variety of scientific and engineering projects

• Coupled atmosphere-ocean models

• Future space vehicle design

• Large-scale dark matter halos and galaxy evolution

Leveraging InfiniBand backward and future compatibility

System Example: NASA Ames Research Center Pleiades

High-Resolution Climate Simulations

High-Performance Designed 100Gb/s Interconnect Solutions

Transceivers

Active Optical and Copper Cables

(10 / 25 / 40 / 50 / 56 / 100Gb/s)VCSELs, Silicon Photonics and Copper

36 EDR (100Gb/s) Ports, <90ns Latency

Throughput of 7.2Tb/s

7.02 Billion msg/sec (195M msg/sec/port)

100Gb/s Adapter, 0.7us latency

150 million messages per second

(10 / 25 / 40 / 50 / 56 / 100Gb/s)

32 100GbE Ports, 64 25/50GbE Ports

(10 / 25 / 40 / 50 / 100GbE)

Throughput of 6.4Tb/s

Leading Supplier of End-to-End Interconnect Solutions

Software and ServicesICs Switches/GatewaysAdapter Cards Cables/ModulesMetro / WAN

StoreAnalyzeEnabling the Use of Data

At the Speeds of 10, 25, 40, 50, 56 and 100 Gigabit per Second

Comprehensive End-to-End InfiniBand and Ethernet Portfolio

The Performance Power of EDR 100G InfiniBand

28%-63% Increase in Overall

System Performance

End-to-End Interconnect Solutions for All Platforms

Highest Performance and Scalability for

X86, Power, GPU, ARM and FPGA-based Compute and Storage Platforms

10, 20, 25, 40, 50, 56 and 100Gb/s Speeds

X86Open

POWERGPU ARM FPGA

Smart Interconnect to Unleash The Power of All Compute Architectures

Technology Roadmap – One-Generation Lead over the Competition

2000 202020102005

20Gbs 40Gbs 56Gbs 100Gbs

“Roadrunner”Mellanox Connected

1st3rd

TOP500 2003Virginia Tech (Apple)

200Gbs

Terascale Petascale Exascale

Mellanox

Maximize Performance via Accelerator and GPU Offloads

GPUDirect RDMA Technology

GPUs are Everywhere!

GPUDirect RDMA / Sync

GPUChip

GPUMemory

System

Memory

Eliminates CPU bandwidth and latency bottlenecks

Uses remote direct memory access (RDMA) transfers between GPUs

Resulting in significantly improved MPI efficiency between GPUs in remote nodes

Based on PCIe PeerDirect technology

GPUDirect™ RDMA (GPUDirect 3.0)

With GPUDirect™ RDMA

Using PeerDirect™

GPU-GPU Internode MPI Latency

Performance of MPI with GPUDirect RDMA

88% Lower Latency

GPU-GPU Internode MPI Bandwidth

10X Increase in Throughput

Source: Prof. DK Panda

2.18 usec

Mellanox GPUDirect RDMA Performance Advantage

HOOMD-blue is a general-purpose Molecular Dynamics simulation code accelerated on GPUs

GPUDirect RDMA allows direct peer to peer GPU communications over InfiniBand• Unlocks performance between GPU and InfiniBand

• This provides a significant decrease in GPU-GPU communication latency

• Provides complete CPU offload from all GPU communications across the network

2X Application

Performance!

GPUDirect Sync (GPUDirect 4.0)

GPUDirect RDMA (3.0) – direct data path between the GPU and Mellanox interconnect

• Control path still uses the CPU

- CPU prepares and queues communication tasks on GPU

- GPU triggers communication on HCA

- Mellanox HCA directly accesses GPU memory

GPUDirect Sync (GPUDirect 4.0)

• Both data path and control path go directly

between the GPU and the Mellanox interconnect

Number of nodes/GPUs

2D stencil benchmark

RDMA only RDMA+PeerSync

27% faster 23% faster

Maximum Performance

For GPU Clusters

Remote GPU Access through rCUDA

GPU servers GPU as a Service

rCUDA daemon

Network InterfaceCUDA

Driver + runtimeNetwork Interface

rCUDA library

Application

Client Side Server Side

Application

Driver + runtime

CUDA Application

rCUDA provides remote access from

every node to any GPU in the system

CPUVGPU

GPUGPUGPUGPUGPUGPUGPUGPUGPUGPUGPU

All HPC Clouds to Use Mellanox Solutions

HPC Clouds

Virtualization for HPC: Mellanox SRIOV

HPC Clouds – Performance Demands Mellanox Solutions

San Diego Supercomputing Center “Comet” System (2015) to

Leverage Mellanox InfiniBand to Build HPC Cloud

Thank You

Interconnect Your Future HPC... · X86, Power, GPU, ARM and FPGA-based Compute and Storage...

Documents

Transcript of Interconnect Your Future HPC... · X86, Power, GPU, ARM and FPGA-based Compute and Storage...

Low Latency Video Analytics Servers · Computing • X86, OpenPower, ARM • GPU and FPGA assist • Push Video encode/decode, streaming, rendering and analytics into the network

Interconnect Your Future - HPC-AI Advisory Council · 2020-01-14 · 5 10 15 20 25 1 4 16 64 256 1K 4K Message Size (bytes) s) GPU-GPU Internode MPI Latency r 67 % 5.49 usec Performance

GPU Architectures - University of Washingtoncourses.cs.washington.edu/courses/cse471/13sp/lectures/GPUs... · Architecture Multicore CPUs x86 ... AMD Graphics Core Next GPU ARCHITECTURES:

X86 Architecture

Harmful x86

X86 Assembly

Introduction to GPU Programming - Home | National … to GPU Programming Volodymyr (Vlad) ... 512-bit memory interconnect ... size_t nbytes, enum

Reference Implementation Software TZMP-1connect.linaro.org.s3.amazonaws.com › hkg18 › presentations › hkg1… · cci – 5xx interconnect gpu nic-400 interconnect vpu display

ACCELERATED COMPUTING & DEEP LEARNING€¦ · Tesla Accelerated Computing Platform GPU Accelerators Interconnect System ... Programming Languages Infrastructure Management System

x86 Memory

Supporting x86 64 Address Translation for 100s of GPU Lanes · Supporting x86-64 Address Translation for 100s of GPU Lanes Jason Power Mark D. Hill David A. Wood Department of Computer

Demystifying Memory Models Across the Computing Stackcheck.cs.princeton.edu/tutorial_slides/ISCA MCM Tutorial v1.1intro.pdfC11/ C++11 Cuda OpenCL x86 CPU ARM CPU Power CPU Nvidia GPU

x86, Assembler

1 Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI ... · 1 Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect Ang Li, Shuaiwen Leon Song, Jieyang

ConFlux: An Ecosystem for Data-enabled Computational Physics · Spark, Hadoop, GraphLab, BlinkDB, DBMS, ... 4TB RAM Pascal GPU + NVLink Throughput Big Data Interconnect (100Gbps speed)

Community Feedback Vulkan SDK Update and€¦ · - X86 vs. x64 differences - Cross OS portability (i.e. capture on windows, replay on linux). - Cross vendor GPU support (capture on

Inter-Module Interconnect Strategy for System on … Interconnect Strategy for System on Chip ... @ic.eecs.berkeley.edu Abstract Interconnect effects that negatively ... Interconnect

Programming GPUs with OpenACC · GPU Hackathon 2017-OpenACC CPU-GPU memory model 3 PCIe Interconnect 16X -8GB/s (gen 2) and 15.75GB/s (gen 3), very thin pipe! Kepler K40 2,880 cuda

High Performance RESEARCH COMPUTING 2300 SUPPORTS …€¦ · dual-GPU accelerators, 16 Intel Knights Landing processor, and an Intel Omni-Path Architecture (OPA) interconnect. ADA

Moving to IBM PureFlex System x86-to-x86 Migration