Upload
mictc
View
1.069
Download
2
Embed Size (px)
Citation preview
Microsoft Innovation Center for Technical
Computing
MICROSOFT AZURE IN HPC SCENARIOS
Lukasz Miroslaw, Ph.D.
18.11.2015, MICROSOFT SWITZERLAND
Challenges
2
57 % of users are dissatisfied with their desktop
computing capacity*
* Source: US Council of Competitiveness: http://www.compete.org, theubercloud.com
Computing: too slow
Memory: too small
Fig. Sometimes solving a problem with IT is hard.
Low Cost of running in the cloud
Cost Model assumes that the hardware makes 7% of Total Costs
4
Fig. Cost of a GFLOP in U.S. Dollars on different Microsoft Azure
nodes and a private HSR cluster.
Agenda
6
Use Case #1: Remote physical simulations with external partners.
Use Case #2: Scale out physical simulations in the cloud.
Use Case #3: Stellar Classification, Prediction of Energy Efficiency in
buildings
Conclusions.
Agenda
7
Use Case #1: Remote physical simulations with external partners. Azure IaaS, Remote App, Azure Batch
Use Case #2: Scale out physical simulations in the cloud.SimplyHPC, HPC Pack
Use Case #3: Stellar Classification, Prediction of Energy Efficiency in
buildings
AzureML
Conclusions.
What is Computational Fluid Dynamics?
CFD is the science to simulate fluid flow, heat and mass transfer and
chemical reactions
8
What is Computational Fluid Dynamics?
Airflow simulation around sky-diving Santa Claus.
9
* Source: Desktop Engineering
Use Case #1: Collaborative Simulations of Electrical Arcs
10
Use Case #1: Collaborative Simulations of Electrical Arcs
Goal #1: Develop a Cloud-based algorithm for electrical arc simulation
Microsoft Azure Research Award in 2014
Contact: Kenji Takeda (Microsoft Research)
Goal #2: Provide simulation tool to partners in Brasil and Deutschland
Ongoing collaborations
Streamer International (CTI Project)
Panasonic
Fraunhofer SCAI
WEG
11
1st Use case: Instant ANSYS
12
VM:
D14 with 16 core
CPU, 112 GB RAM,
Windows Server
2012
MpCCI, ANSYS
preinstalled
Storage: locally
redundant,
automatically
scallable
License Server (LS)
on A0 in Germany
Customer
VM
LS
INSTANT ANSYS
13
No Installation. No configuration. No up-front costs.
Access to powerful VMs with ANSYS already preinstalled
and preconfigured.
Access to redundant and highly available storage.
Disaster Recovery and 99.5% SLA.
Connection to on-premise infrastructure with IPSec VPN.
2nd Use case: Linux VM
15
The UberCloud: Making Technical
Computing available in the Cloud
UberCloud Community:
+2500 companies and
individuals:
+60 cloud providers,
+80 software providers,
several hundred consulting
firms and individual experts.
OpenFOAM added to Azure
Marketplace
Docker containerization
www.ubercloud.com
2nd Use case: Linux VM
16
DEMO
The compute environment you ordered is now
ready.
Access your compute environment via remote
desktop connection (Chrome 8+, Firefox 7+,
Opera 11+, IE 9+)
Launch
Your password for remote desktop access is:
TN1b39pv4Djw
Azure RemoteApp
17
Deliver apps from the cloud, cost-
effectively
Simplify your infrastructure
Run Windows apps anywhere
Centralize your apps, help secure your data
Costs
19
VM with 16 cores and 56 GB RAM costs 2.11 CHF / hour (D14)
1 TB of Storage costs 30 CHF / month
RemoteApp starting price: $10 / user / month (40h included)
Online Calculator
Azure in Education
Faculty will receive a 12 month,
$250/month account
Students will receive a 6 month,
$100/month account
Short Summary
20
+ Powerful VMs that can be started/stopped on-demand increase
the productivity in our group.
+ Virtual images with OS and different software version to avoid
problems with backward compatibility.
+ Students and team members can manage their own VMs and
reduce the costs of support.
- Storage File Service can be easily mapped to a drive on the VM
but not on premises.
- Only a single user can access one VM.
SimplyHPC: Light-weight Cloud Orchestrator for MS Azure
What is SimplyHPC?
Framework
23
SimplyHPC:
1) Distributed framework for
Microsoft Azure,
2) Set of PowerShell scripts.
Performance and Scalability
Example #1: Solving linear systems with PETSc and HPCG
25
Fig. Performance in GFlops of PETSc solving ruep (right) matrix
system and HPCG Benchmark (left) on different Microsoft Azure
nodes and a private HSR cluster.
Performance and Scalability
27
Fig. Strong scaling of ANSYS CFX of the compressor (11 mln nodes).
Example #2: ANSYS CFX
Azure Batch
28
Batch is a managed service for batch
processing or batch computing - running a
large volume of similar tasks to get some
desired result.
Short Summary
SimplyHPC: a framework to simplify cluster deployment and
job submission.
Set of light-weight PowerShell scripts to submit, execute and
monitor multi-threaded jobs on Windows Azure.
Easy to use. No cloud-related knowledge necessary.
Run the jobs from command line and download the results
directly to your Azure Storage.
Up to 9x faster than native MS HPC Pack scripts.
Available at https://github.com/vbaros/SimplyHPC
29
L Miroslaw, V Baros, M Pantic, H Nordborg, Unified Cloud Orchestration
Framework for Elastic High Performance Computing on Microsoft Azure,
NAFEMS World Congress 2015
Short Summary
30
+ Scaling properties of Microsoft Azure is comparable to the on-premises
cluster.
HSR Cluster: 7.3 days (176 hours), limited availability.
Microsoft Azure: 4.9 days (118 hours), ca. 50% faster, 100% availability.
+ Dynamic scaling (up- / downscaling) and instant access to the newest
hardware reduces the costs.
+ (Un)limited computing at competitive price.
Cluster composed of 32 x A8 nodes (=256 cores) costs
32 x 2.11 CHF/h = ca. 68 CHF/h
- Upscaling > 100 cores should be planned in advance.
Microsoft Azure Machine Learning Studio
Three types of knowlege:
Know-What (facts)
Know-How (processes)
Know-Why (reasons)
31
Image credit: Univ. Hamburg
AzureML Studio
Key goals of Machine Learning:
Prediction
Classification
Clustering
Collaborative Filtering
32
Image credits: OpenCV, Snipview, Stanford
AzureML: Stellar Classification
Classification Challenge:
HYG database* is a compilation of
of stellar data from three main
catalogues.
Contains ca. 120k stars, 37
spectral characteristics.
2D classification scheme based on
temperature (color index) and
brightness (absolute magnitude).
Data is incomplete and may
contain a few misclassifictions.
Prediction Engine developed in
AzureML
33
Credits: Michael Pantic (HSR)* http://www.astronexus.com/hyg
AzureML Example: Heating Load Prognosis
34
Image credit: SAB Magazine
Input:
- Roof area
- Overall hight
- Glazing area
- Surface area
- ...
Output:
- Heating load
prediction
AzureML Workflow
35
Machine Learning Workflow
1. Hypothesis
2. Data Preparation
3. Model
4. Test
5. Evaluate
A. Tsanas, A. Xifara: 'Accurate quantitative estimation of energy performance of
residential buildings using statistical machine learning tools', Energy and Buildings,
Vol. 49, pp. 560-567, 201
- 8 physical characteristics
from 768 buildings
- Goal: predict buldings’
heating load and cooling
load
- Architects need to compare
several building designs
before selecting the final
approach
AzureML: Short Summary
37
Very fast prototyping. Load the system with data, test different
Machine Learning methods.
Platform for Internet of Things: Event Hubs, Stream Analytics.
Share the models & results.
Deploy web services fast.
Develop own methods in Python and R Statistics.