成本節約之道：加速設計週期 x 大規模運行高效能運算 (HPC) 工作負載 (Level: 300)

© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
© 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.© 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Jhen-Wei Huang (黃振維)
Solutions Architect, Amazon Web Services
High Performance Computing on AWS

© 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Agenda
 Overview of AWS Infrastructure
 Why HPC on AWS
 HPC Solution Components
 Cost Optimization
 Performance Considerations

AWS Global Infrastructure

Over 100 Global CloudFront PoPs
AWS Global Infrastructure
Regions
Amazon Global
Network
• Redundant 100 GbE network
• Redundant private capacity
between all Regions except China

AWS Global Infrastructure Region & Number of Availability Zones
US West EU
Oregon (3) Ireland (3)
Northern California (3) Frankfurt (3)
London (3)
US East Paris (3)
N. Virginia (6), Ohio (3)
Asia Pacific
Canada Singapore (3)
Central (2) Sydney (3), Tokyo (4),
Osaka-Local (1)
Seoul (2), Mumbai (2)
AWS GovCloud
US-West (3)
China
South America Beijing (2)
São Paulo (3) Ningxia (2)
55 Availability Zones within 18 geographic
Regions and 1 Local Region around the world
Announced Regions
Bahrain, Hong Kong, SAR(China), AWS
GovCloud (US-East)

Why HPC on AWS?

Amazon Running HPC Workloads Everyday
 Logistics
 Machine learning
 Data center, network, and server design
 Consumer product design
 Robotics
 Semiconductor design
 Retail and financial analytics

Why HPC on AWS
Faster Time to
Results
Better ROI
 Virtually unlimited infrastructure
enabling scaling and agility not
attainable on-premises
 Flexible configuration options
quickly iterate resource selection
and ensure cost optimization
 Increased collaboration with
secure access to clusters around
the world

Why HPC on AWS – Multiple Clusters
$ qsub –q monolith iwait.sh
$ qsub dev.sh
$ qsub prod.sh
$ qsub critical.sh
$ qsub bigrun.sh
On-Prem
Launch clusters by group, user,
application – no more waiting!

Transcoding and
Encoding
Monte Carlo
Simulations
Computational
Chemistry
Government and
Educational Research
Modeling and
Simulation
Genome Processing
Popular HPC Workloads on AWS

HPC Workload Types
Tightly Coupled
Parallel
Computing
Loosely Coupled
Parallel
Computing
Accelerated
Computing
Visualization and
Interpretation
High Performance
Data Storage and
Analytics
 Similar
instance types,
fixed size
clusters of EC2
instances
 Network
intensive
 Customers
price sensitive
 High
utilization
 Not typically
resilient to
interruptions
 Scalable,
flexible
infrastructure
 Workloads are
also easily
interruptible
 Data Intensive
 Typically
massively
parallel
application
 Need compute
optimized
GPUs or FPGAs
 Workloads run
on graphics-
optimized
GPUs
 Need
additional
managed
services like
Workspaces or
AppStream 2.0
 Workloads
require
moving
customer data
to AWS
 Value creation
based on
innovative
analytics
strategies like
AI/ML/DL

Industry Verticals and common HPC Workloads
 Computational
Chemistry
 Genomics
 Proteomics
 Bioinformatics
 Neuroimaging
 Clinical Trials
Simulations
 Molecular
Dynamics
 RNA Sequencing
 Risk analysis /
modeling
 Regulatory
compliance
 Monte Carlo
simulation
 Actuarial Grid
 High Frequency
Trading
 Bitcoin / Block
chain
 Weather
Simulation
 Reservoir
Simulation
 Geographical
Information
Systems
 Operations,
management,
and analysis
 Electronic
design
automation
 Computational
fluid dynamics
 Engineering
Simulations
 BIM
 FEA
 Rendering
 Content
Creation
 Post production
Life Sciences Financial Services
Energy &
Geosciences
Manufacturing
Media and
Entertainment

Global-scale grids for research
Large Hadron Collider (LHC)

Global-scale grids for research
Best-practices using Spot: diversify computing with many instance types,
multiple AZs, multiple regions, and with stateless architectures

1.1M vCPUs for machine learning

HPC grids in financial services
U s i n g G P U A c c e l e r a t i o n The challenge
Spinning up large numbers of GPUs quickly and
inexpensively to meet ABSI’s customers financial
modeling and reporting needs
ABSI uses proprietary algorithms (Monte Carlo
simulations) running millions of times
The solution
ABSI moved its infrastructure to AWS and deprecated its
co-located data center
ABSI built a front end on AWS for its processing
solution, automatically running GPU instances on
Amazon EC2 using Amazon EBS in an Amazon VPC for
security.
The result
Can be as much as 500 times more efficient in terms of
performance per dollar for some clients
“Using AWS helps us reduce a 10-
day process to 10 minutes. That’s
transformative: it broadens our
ability to discover.”
–Peter Phillips
Managing Director, Aon Benfield Securities

HPC in design and manufacturing
Applications for engineering:
 Molecular dynamics, CAD, CAE, EDA
 Collaboration tools for engineering
 Big data for manufacturing yield analysis
Running drive-head
simulations at scale:
Millions of parallel parameter
sweeps, running months of
simulations in just hours
Over 85,000 Intel cores
running at peak, using Spot
Instances

Tightly coupled HPC—weather

Fluid dynamics—Ansys Fluent
 C4.8xlarge instance type
 140M cell model
 F1 car CFD benchmark

HPC Solution Components

Important enablers in HPC
 Compute performance—CPUs, GPUs, FPGAs
 Memory performance—high RAM requirements in many applications
 Network performance—throughput, latency, and consistency
 Storage performance—including shared filesystems
 Automation and cluster/job management
 Graphics for pre- and post-processing
…and SCALE

HPC Solutions Storage
EBS EFS
S3
Networking
Enhanced
Networking
Placement
Groups
Automation &
Orchestration
AWS Batch
CfnCluster
NICE EnginFrame
Visualization
NICE DCV
AppStream 2.0
Compute
EC2 Instances
(Compute and Accelerated)
EC2 Spot
Auto Scaling

Amazon EC2 Instances
General
purpose
Dense
storage
Compute
optimized
FPGA
GPU
Compute
Storage
optimized
Graphics
intensive
Memory
optimized
High
I/O P2M4 D2 X1 G2T2 R4I3C5 F1M5 P3H1 EC2 Bare MetalG3T2 Unlimited X1eI2C4
High
I/O
General
purpose
burstable Direct access to
physical server
resources
Optimize the price/performance of your HPC Workloads with the
widest range of compute instances

C5 Instances - Intel XEON Scalable Processor
 Intel Skylake @ 3.0
GHz (turbo to 3.5 GHz)
 Supports AVX512
 C-state controls
 Nitro System, a
combination of
dedicated hardware and
lightweight hypervisor
 Up to 25-Gbps network
AVX 512
72 vCPUs
“Skylake”
144-GiB memory
C5
12 Gbps to EBS
2X vCPUs
3X throughput
2.4X memory
C4
36 vCPUs
“Haswell”
4 Gbps to EBS
60-GiB memory

Xilinx
UltraScale+
FPGANVIDIA GPU
P2/P3: GPU-accelerated computing
 Enabling a high degree of parallelism
– each GPU has thousands of cores
 Consistent, well documented set of
APIs (CUDA, OpenACC, OpenCL)
 Supported by a wide variety of ISVs
and open-source frameworks
F1: FPGA-accelerated computing
 Massively parallel – each FPGA includes
millions of parallel system logic cells
 Flexible – no fixed instruction set, can
implement wide or narrow datapaths
 Programmable using available, cloud-
based FPGA development tools
GPU and FPGA for Accelerated Computing

Deep learning on GPU
MXNet training on EC2 P2
instances:
 Training of a popular image
analysis algorithm, Inception v3,
using MXNet and running on P2
instances
 Scaling efficiency of 85%

FPGA use-cases and F1 partners

Network Performance
AWS Proprietary Network, 10 Gbps & 25 Gbps
 Highest performance in largest EC2 instance sizes
 Full bisection bandwidth in Placement Groups, with no network oversubscription
Enhanced Networking
 Over 1M PPS performance, reduced instance-to-instance latencies, more
consistent network performance
Amazon EC2 to Amazon S3
 Traffic to and from Amazon S3 can now take advantage of up to 25 Gbps of
bandwidth

Instance sizes: R4 example
R4 instances are optimized for
memory-intensive applications
 Xeon E5-2686 v4 processors
 DDR4 Memory
 Enhanced Networking, up to
25 Gb throughput

Amazon EFS
File
Amazon EBS
Amazon EC2
Instance Store
Block
Amazon
S3 / S3-IA
Amazon Glacier
Object
Data Transfer
AWS Direct
Connect
ISV
Connectors
Amazon
Kinesis Data
Firehose
AWS Storage
Gateway
S3 Transfer
Acceleration
AWS Storage is a Platform
AWS
Snowball
Amazon
CloudFront
Internet/
VPN

Storage Classes and Tiering on Amazon S3
Standard
• Primary data
• Big data analytics
• Small objects
• Temporary scratch space
• Archive data
• Deep/offline archives
• Tape vaulting replacement
• WORM-compliant data
• File sync and share
• Active archive
• Enterprise backup
• Media transcoding
• Georedundancy/DR
Standard - Infrequent Access Amazon Glacier

Block Storage
Two Block Storage options for EC2 Instances: Amazon EBS and Instance Store
EC2 Instance
/dev/xvda
/dev/xvdb
/dev/xvdc
Block Device Mapping Instance Store
ephemeral0
ephemeral1
vol-xxxxxxxx
vol-xxxxxxxx
/dev/xvdd
EBS Volumes

File Systems on AWS
 Amazon Elastic File System (Amazon EFS)
 Distributed across multiple AZs
 Petabyte-scale
 Easy to bring up, no management
 Build your own NFS
 Use for a POC
 Ephemeral data (i3.*)
 Parallel file systems
 Build your own or use APN solutions

High-performance NFS on AWS
 EC2+EBS is the most performant method of
creating scale-up file servers on AWS
 Build your own NFS or CIFS implementation
or use a partner solution
 EC2 instances as fileservers, using EBS for
block storage—tuned for application needs
 Single fileserver performance up to 25 Gb/s
over the network

Amazon S3
Secure, durable, highly
scalable object
storage. Fast access,
low cost.
For long-term durable
storage of data, in a
readily accessible
get/put access format.
Primary durable and
scalable storage for
critical data
Amazon Glacier
Secure, durable, long
term, highly cost-
effective object
storage.
For long-term storage
and archival of data
that is infrequently
accessed.
Use for long-term,
lower-cost archival of
critical data
EBS+EC2
Create a single-AZ,
shared file system
using EC2 and EBS,
with third-party or
open source software
(ZFS, Weka.io, Avere,
Intel Lustre, etc.).
For near-line storage
of files optimized for
high IOPS.
Use for high-IOPS,
temporary working
storage
Optimize HPC storage
Amazon EFS
Highly available,
multi-AZ, fully
managed network-
attached elastic file
system.
For near-line, highly-
available storage of
files in a traditional
NFS format (NFSv4).
Use for read-often,
temporary working
storage

Data transfer
HPC Data Flow on AWS Storage
Corporate data center
Amazon
Glacier
Amazon S3
AWS Direct
Connect
ISV
Connectors
Storage
Gateway
AWS
Snowball
Internet/VPN
Ingress
Egress
Lifecycle
EC2 Instance
EBS
Instance
Store
Object, block, file storage
Kinesis Data
Firehose
S3 Transfer
Acceleration
Amazon
CloudFront
Other Shared File
System
EFS
25 Gbps to S3

Automation and Batch Processing

Traditional Job Schedulers Integrate Easily
Bring your scheduler to AWS, or build your own
 IBM Platform LSF
 Univa Grid Engine
 Altair PBS Pro
 SLURM
 Design your own using AWS services
 Do you actually need a scheduler?

HPC Automation and Orchestration
Choose from several options to adapt your workloads
 CfnCluster
 AWS Batch
 AWS-NICE DCV and EnginFrame
 Build your own AWS CloudFormation templates
 ISV offerings on AWS Marketplace or use an SI

HPC automation with CfnCluster
CfnCluster simplifies
deployment of HPC in the cloud,
including integrating with
popular HPC schedulers
Built on AWS CloudFormation,
easy to modify to meet specific
application or project
requirements

Fully managed
No software to install or
servers to manage. AWS
Batch provisions, manages,
and scales your
infrastructure.
Integrated with AWS
Natively integrated with the
AWS Platform, AWS Batch
jobs can easily and securely
interact with services such as
Amazon S3, DynamoDB, and
Amazon Rekognition.
Cost-optimized
resource provisioning
AWS Batch
automatically provisions
compute resources
tailored to the needs of
your jobs using Amazon
EC2 and EC2 Spot.
AWS Batch for HPC workloads
Focus on your applications and results!

HPC Architecture on AWS

HPC Architecture on AWS
Corporate data center
Availability Zone
Auto Scaling group
Parallel
FS
Local
NFS
S3
Data
ingress/egress
Amazon
EFS
 Three file systems: Amazon
EFS, Local NFS, and Parallel
FS
 Snapshot of Amazon EBS to
Amazon S3
 Data tiering FS to Amazon S3
 Auto Scaling allows for scaling
when needed
Master instance
$ qsub job.sh
EBS

Hybrid HPC
U s i n g A m a z o n
E C 2 S y s t e m s M a n a g e r
Capabilities
Run
Command
State
Manager
Inventory Maintenance
Window
Patch
Manager
Automation Parameter
Store
Documents
AWS cloud
corporate data
center
IT admin, DevOps engineer
Role-based access control
 Manage thousands of Windows and Linux nodes
running on Amazon EC2 or on premises
 Control user actions and scope with secure,
granular access control
 Safely execute changes with rate control to
reduce blast radius
 Audit every user action with change tracking

Graphics for HPC applications

Desktop application streaming
with Amazon AppStream 2.0
Stream desktop applications securely
to any web browser
Pay as you go Scale globally
Secure apps and dataRun desktop apps
in a web browser

AppStream 2.0 graphics support
 Multiple Instance types—including General Purpose,
Compute Optimized, Memory Optimized, Graphics
Design, Graphics Pro, and Graphics Desktop
 Always-On or On-Demand pricing models
 Support for OpenGL, DirectX, OpenCL and CUDA

Cost Optimization

EC2 Purchasing Options
On-Demand
Pay for compute capacity by
the second with no long-
term commitments
Spiky workloads, to define
needs
Reserved
Make a 1- or 3-year commitment
and receive a significant discount
off On-Demand prices
Committed, steady-state usage
Spot
Spare EC2 capacity at savings of
up to 90% off On-Demand prices
Fault-tolerant, dev/test, time-
flexible, stateless workloads
Per Second Billing for EC2 Linux instances & EBS volumes

Cost Optimization
Weather Forecasting and Modeling
On demand
Spot
Reserved
instances
Forecasting
00z, 06z, 12z, 18z
Climate
modeling
Weather
events
Daily forecasts
Climate
modeling
Hurricane

Performance considerations

f o r t i g h t l y c o u p l e d c l u s t e r w o r k l o a d s
Test using real-world examples
 Use large cases for testing: do
not benchmark scalability using
only small examples
Domain decomposition
 Choose number of cells per core
for either per-core efficiency or
for faster results
MPI libraries
 Test with Intel MPI and
OpenMPI 3.0, and make use of
available tunings
Network
 Use a placement group
 Enable enhanced networking
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
0.0
500.0
1000.0
1500.0
2000.0
2500.0
0 500 1000 1500 2000 2500 3000 3500 4000
Time(S)
Scale-Up
Cores
WRF 2.5 km CONUS Benchmark
Scale-Up time

f o r a l l H P C w o r k l o a d s
OS version
 Use Amazon Linux or an
updated 3.10+ kernel–4.0+ if
using NVME on F1 or I3
Processor states
 Use P-states to reduce
processor variability
Instance types
 C5, C4, M4, R4 are the best
choices today—but always test
with the latest EC2 instances
Hyper-threading and affinity
 Test with Hyper-Threading (HT)
on and off—usually off is best,
but not always
 Use CPU affinity to pin threads
to CPU cores when HT is off

Thank you, and how can I help you
run HPC workloads on AWS?
aws.amazon.com/hpc

成本節約之道：加速設計週期 x 大規模運行高效能運算 (HPC) 工作負載 (Level: 300)

Recomendados

Recomendados

Más contenido relacionado

La actualidad más candente

La actualidad más candente (20)

Similar a 成本節約之道：加速設計週期 x 大規模運行高效能運算 (HPC) 工作負載 (Level: 300)

Similar a 成本節約之道：加速設計週期 x 大規模運行高效能運算 (HPC) 工作負載 (Level: 300) (20)

Más de Amazon Web Services

Más de Amazon Web Services (20)

成本節約之道：加速設計週期 x 大規模運行高效能運算 (HPC) 工作負載 (Level: 300)