Heavy duty Abaqus structural analysis using HPC in the cloud.
Frank Ding presents challenges, solutions, lessons Learned, and recommendations.
Visit HPCExperiment.com to learn how you can start using HPC in the Cloud.
HPC as a Service: Paving the Way with Heavy Duty Abaqus Analysis
1. Paving the way to HPC as a Service
Heavy Duty Abaqus Structural Analysis Using HPC in the Cloud –
Challenges, Solutions, Lessons Learned, & Recommendations
Frank Ding
May 16, 2013
2. What is the UberCloud HPC Experiment?
Started in mid-2012 as a voluntary effort
Demonstrate the potential of HPC in the Cloud
Uncover and overcome the obstacles
Now with over 450 participants Round 3 is in progress
Approaching 80 teams as of today
Please submit your project ideas at the end of this webinar!
2
3. About the presenter
Frank Ding, Engineering Analysis & Computing
Manager, Simpson Strong-Tie, HPC Experiment
Advocate,An early adopter of HPC for simulation-driven
design, a huge fan of Linux-based cluster computing
Thanks to members of the HPC ExperimentTeam #47
Matt Dunbar, SIMULIA 3DS
Steve Hebert & Rob Sherrard, Nimbix
Sharan Kalwani, DataSwing
Antonio Arena & Cynthia Underwood, NICE Software
Dennis Nagy, BeyondCAE
3
4. Simpson Strong-Tie
A leader in structural systems research, testing and innovation
Product Lines
Connectors for light frame construction
Fasteners and fastening systems
Lateral systems
Anchoring systems
Fiber reinforcing materials
4
6. Why HPC in the Cloud?
Current HPC Cluster – 4-node 32 cores total
Nehalem-based Xeon
InfiniBand DDR
HPC in the Cloud
Lack of in-house HPC resources
Capacity surge – large jobs speedup
Capacity overflow jobs – large number of jobs
Project tested in Round 2 HPC Experiment
Concrete anchor bolt tension capacity
1.9 millions of DOF’s
11.5 hours runtime (32-core)
6
7. Cloud-Based HPCWorkflow
ProjectWorkflow
Pre-processing on the local workstation at the user end
Abaqus input file loaded to the data staging point thru SFTP
Abaqus job submitted to the compute cloud thru Nimbix web portal
Job monitoring thru Nimbix dashboard plus email notification of job status
Post-processing using remote visualization tool NICE DCV
7
8. Barriers and Challenges
Data movement limited by internet bandwidth
Use remote visualization for post-processing
End user side internet bandwidth issue
Team member time schedule (voluntary effort)
Team member expertise gap
Meet the project deadline
8
9. Benefits
HPC in the Cloud solutions require multi-vendor support
ISV – 3DS SIMULIA provided Abaqus license
Cloud infrastructure and service provider – Nimbix HPC Cloud for CAE
Remote visualization – NICE Desktop CloudVisualization(DCV)
End-user applications
HPC Experiment provides a collaboration platform
Form a team based on end-user application requirement
Basecamp.com to support team communication
Third-party solution providers invited to the team if needed
9
10. Lessons Learned & Recommendations
End point internet bandwidth variability is the top barrier
Some workflow details have been identified to improve end user
experience
Service provider is recommended to provide connection bandwidth
testing tool to the end users
Remote visualization using NICE DCV is a good platform if a
stable/consistent internet bandwidth is available. Results accessible
anywhere when needed.
HPC Experiment is a great platform to test your project on the Cloud
HPC
10
11. Ready for your project!
Round 3 is in progress and will conclude at the end of June
Anyone can create a project
And it is not too late for Round 3 (and beyond)
Please fill out the form at the end of this webinar
11
12. Q&A session notes
What bandwidth is needed in your opinion?
FD: Roughly a consistent bandwidth over 5Mbps is required
for a smooth response for dynamic 3D model manipulation.
What is the largest memory node/core on the cloud
forAbaquspurposes?
FD: I use 4GB/core for my local HPC, but I don’t know Nimbix
Compute Cloud’s configuration, but I would recommend the
same.
13. Q&A session notes
Do you use .pbs to submit in batch?
FD: No, I use Nimbix Compute Cloud job submission web portal,
which talks to theTorque load manager.
What sort of interconnect types did you use / try?
FD:We used GigE in the Round 1 and InfiniBand in Round 2.
InfiniBand performs better.
Can you talk about the Memory required for your project?
FD:The job was solved by Abaqus/Explicit, which does not have
intensive memory requirement like Abaqus/Standard does. Usually
I go with 4GB/core.
14. Q&A session notes
You mentioned that Internet speed (or lack of) was a major
factor.What was the maximum internet speed available to
you for the project?
FD:The maximum I ever tested was 10 Mbps on a 20 Mbps
pipe.
Is the NICE visualization using a secure channel (encryption)?
FD:Yes, encryption using the standard AES algorithm (128 or
256-bit)