Porting and scaling engineering applications in the cloud Wolfgang Gentzsch, UberCloud
Porting and scaling engineering applications in the cloud
Wolfgang Gentzsch, UberCloud
UberCloud Community & Industry Validation
3000+Community Members
200+HPC Cloud Experiments
Lessons from the experiments go directly
into the product
Established expertise and thought leadership
40+Marketplace Stores
Online access to
CAE technology
and expertise
50+HPC Containers
ANSYS, CD-adapco,
COMSOL,GROMACS,
NUMECA, SIMULIA, etc
Porting scenarios
Your application
Your container
Your servers
Any HPC Cloud
Your Workstation(s)
Benefits of Public Clouds
• More (infinite) computing
• No upfront Cap-Ex investment
• On demand, pay per use, at your fingertips
• Scaling resources dynamically, up and down
• Always the latest hardware and software
• No long procurement, nor acquisition cost, nor high TCO
• No need for expensive on-premise infrastructure & experts
• Choice, with multiple providers
10 Guidelines for Porting Your Application to the Cloud
1. Evaluate benefits of cloud vs on-premise for your specific scenario:On-premise or bursting or hosting or hybrid cloud? More jobs (more parameters) or larger models or ‘more‘ physics (FSI)
2. Look for existing cloud benchmarks: kernels, solvers, similar applications, and consider vendor and in-house benchmarks: on in-house workstation and server vs cloud node and server
3. Cloud providers might help with your benchmarking4. Cloud case studies and articles: contact / consult their authors 5. Different porting scenarios for in-house, open-source, commercial
software
10 Guidelines for Porting Your Application to the Cloud
6. Find out level of scalability of your application software => what scales on your in-house system should scale in the cloud
7. Select HPC cloud provider(s) with: powerful compute nodes, large memory, Infiniband, GPUs,…, user-friendly access, batch and interactive,… Azure, Bull/Atos, CPU 24/7, Nephoscale, Nimbix, R-Systems, Sabelcore, etc.
8. Consider HPC software container for your application: ease of packaging, porting, accessing, using, and scaling, plus maintenance and support
9. Perform a trial / proof of concept first10. Start with your existing model, compare results, increase in small steps
Porting in-house, open source, and commercial software
In-house Software- Pro: flexibility, fully under your control, no licensing, and in-house s/w expertise - Con: proprietary, high dev/maintenance/porting cost, and in-house s/w expertise- Cloud: building your own cloud environment and port yourself
Open Source Software- Pro: community effort, no s/w dev and maintenance, no licensing, - Con: often bugs, no reliable service and support, in-house s/w expertise- Cloud: building your own cloud environment and port yourself or the community
or cloud providers have done it already and offer it as a service
Commercial Software- Pro: stable, reliable, 24/7 support & maintenance, regular updates- Con: not flexible, expensive, limited capabilities (e.g. #cores), vendor lock-in- Cloud: ‚self-porting‘ not possible, either ISV or UberCloud have done it, licensing
often limited, vendor dependent, BYOL only with existing license, no easy upgrade/scale
Looking for the right HPC Cloud ProviderBenchmarking of cloud vendors with High Performance Linpack*
• Azure, AWS, IBM SL, Rackspace, NERSC
• Azure HPC scalability is comparable to an HPC Cluster at scale
• AWS still the first Public Cloud provider people think of, but in HPC it is not
• Other public clouds struggle in HPC (e.g. IBM Softlayer)
Mohammad Mohammadi, Timur Bazhirov, Exabyte Inc. Feb 2017
*Solver for large algebraic equation systems
Team 171:
UberCloud Team 171: Dynamic Study of Frontal Car
Crash with UberCloud ANSYS Container in the Cloud
Simulation times for different numbers of cores for a mesh model with 17K elements. Conclusion: number of elements is too small for higher number of parallel cores!
End-User: Praveen Bhat, Technology Consultant, INDIA. Software Provider: ANSYS in UberCloud Container. Resource Provider: Nephoscale in California
UberCloud Team 180: CFD SpeedIT on GPUs in the Cloud *
Test cases from top left: AeroCar & SolarCar - geometries from 4-ID Network; motorBike - geometry from OpenFOAM tutorial; DrivAer - geometry from Institute of Aerodynamics and Fluid Mechanics at TUM
Scaling OpenFOAM on 2 clusters from Ohio Supercomputing Center: Oakley with Intel Xeon X5650 processors and Ruby with Intel Xeon E5-2670 v2 processors, and on a single GPU (NVIDIA Tesla K40), using SpeedIT Flow
*) Case Study Author – Andrzej Kosior, Vratis Ltd.
Team 181
UberCloud Team 181: Prediction of Barehull KRISO
Containership Resistance with NUMECA in the Cloud
**
*) UberCloud Container on CPU 24/7
Comparing NUMECA FINE/Marine simulation resultson different clouds with corresponding experimental result
UberCloud multi-container multi-node environment
UberCloud Team 175: Parametric Radio Frequency
Heating with COMSOL Multiphysics in the Cloud
This Radio Frequency (RF) heating application simulates dielectric heating of an insulated block, caused by microwaves travelling in an H-bend waveguide.
RF model: a large number of parameters need to be computed. It can be parallelized so that several frequencies and geometric parameters are computed at the same time. This model yields what is called an embarrassingly parallel computation.
Join me at a Live Demo
Aerodynamics of a Motorbike
https://naca4412.theubercloud.net