Date post: | 15-Jul-2015 |
Category: |
Data & Analytics |
Upload: | logilab |
View: | 288 times |
Download: | 2 times |
Collaboration:Aster Pronet – March 27, 2015
Study of the dynamic behavior of a pump with Code_ASTER on Simulagora
Collaboration:
Florent CayréKarina MacoccoFelipe Aguirre
Aster PronetMarch 27, 2015
Collaboration:Aster Pronet – March 27, 2015
… solutions for robust engineering
SICODYN BenchmarkSurrogate modeling
Collaboration:Aster Pronet – March 27, 2015
SICODYN BenchmarkStudy of the pump casing – variables & definition domain
Uncertain parameters are the radius of the connecting screws of the frame with the other elements.
Uniform distributions
Radius type 1 [m] : [0,0088 ; 0,0132]
Radius type 2 [m] : [0,011088 ; 0,016632]
Radius type 3 [m] : [0,011088 ; 0,016632]
● 400 training points and 100 testing points
● 1h30 minutes per simulation with high RAM consumption (10 GB).
● Total simulation time :
– 25 days for the training points (Distributed with Simulagora)
– 6,25 for the testing points (Distributed at Phimeca over 3 cores)
© P
him
eca
En
gine
erin
gLi
bre\
Res
trei
nt
Ph
ime
ca \
Con
fiden
tiel
Collaboration:Aster Pronet – March 27, 2015
Input of the design of experiment
Sobol Sequence
• 400 points for the training dataset.
LHS• 100 points for the testing
dataset. Highlighted in the graph.
© P
him
eca
En
gin
eerin
gLi
bre
\Res
tre
int
Phi
mec
a \
Co
nfid
ent
iel
Collaboration:Aster Pronet – March 27, 2015
Output of the design of experimentFrequencies
© P
him
eca
Eng
inee
ring
Lib
re\R
est
rein
t P
him
eca
\ C
onfid
ent
iel
Collaboration:Aster Pronet – March 27, 2015
MAC matrix of the VGP©
Phi
mec
a E
ngi
neer
ing
Lib
re\R
estr
ein
t P
him
eca
\ C
onfid
entie
l
Collaboration:Aster Pronet – March 27, 2015
Adjustment graph of the VGP©
Ph
imec
a E
ngi
nee
ring
Lib
re\R
estr
ein
t P
him
eca
\ C
onf
ide
ntie
l
Collaboration:Aster Pronet – March 27, 2015
Validation of the VGP©
Ph
ime
ca E
ngin
eerin
gL
ibre
\Re
stre
int
Phi
mec
a \
Con
fide
ntie
l
Collaboration:Aster Pronet – March 27, 2015
Validation of the VGP©
Ph
ime
ca E
ngin
eerin
gL
ibre
\Re
stre
int
Phi
mec
a \
Con
fide
ntie
l
Collaboration:Aster Pronet – March 27, 2015
Simulagora in few words...
Public cloud based
– Computation and storage resources
– Enormous power
– Super-fast computation resources ramp-up
Server
Cloud
Simulagora
Cloud
Simulagora
Browser
WWW
A platform using Web technology to ease grid computation by leveraging cloud resources
HPC in the cloud
Collaboration:Aster Pronet – March 27, 2015
Implementation on Simulagora
Modeling Results analysis
● Modeling and results analysis performed on Phimeca infrastructure● Could have been done on Simulagora too
Single computation setup
Complete study setup and monitoring
Simulagora
Collaboration:Aster Pronet – March 27, 2015
Simulagora - single computation setup
● Transfer input data : using the Web interface
Technicalities
● Data integrity check
● Resume on failure
Collaboration:Aster Pronet – March 27, 2015
Simulagora - single computation setup
● Transfer input data : using the Web interface
Technicalities
● Data integrity check
● Resume on failure
No install
No setup
Collaboration:Aster Pronet – March 27, 2015
● Transfer input data : using the Python API
– Simple config file
– Simple Python script
Technicalities
● Open Source
● Easy install, easy setup
● Web services (no firewall pb)
Simulagora - single computation setup
Collaboration:Aster Pronet – March 27, 2015
● Transfer input data : using the Python API
– Simple config file
– Simple Python script
Technicalities
● Open Source
● Easy install, easy setup
● Web services (no firewall pb)
Simulagora - single computation setup
Fast install
Easy setup
Collaboration:Aster Pronet – March 27, 2015
● Launch an interactive session
Simulagora - single computation setup
Technicalities
● Full root access
● Max session time
● File modifications saved
Collaboration:Aster Pronet – March 27, 2015
Simulagora - single computation setup
Technicalities
● Requires a high-bandwidthconnection
● Full-featured lightweightwindow manager
● Connect to your machine in your browser
Collaboration:Aster Pronet – March 27, 2015
Simulagora - single computation setup
No install
No setup
Technicalities
● Requires a high-bandwidthconnection
● Full-featured lightweightwindow manager
● Connect to your machine in your browser
Collaboration:Aster Pronet – March 27, 2015
Simulagora - single computation setup
● Connect to your machine in text-mode
Technicalities
● Uses the SSH protocol(rarely blocked by firewalls)
● Fast even with low-bandwidth connections
● Requires a simple setup(a bit more on Windows)
Collaboration:Aster Pronet – March 27, 2015
Simulagora - single computation setup
● Connect to your machine in text-mode
Technicalities
● Uses the SSH protocol(rarely blocked by firewalls)
● Fast even with low-bandwidth connections
● Requires a simple setup(a bit more on Windows)
Fast install
Easy setup
Collaboration:Aster Pronet – March 27, 2015
Simulagora – prepare to take-off...
● Write the parametrized launch script and check that :
– it runs smoothly to the end– it produces the expected results
● Prepare the 400 runs
Technicalities
● Also possible to store run parameters into a file per run
● Runs are recorded but notstarted until explicitlyrequested
● Main difficulty is to generatethe parametrized command
Collaboration:Aster Pronet – March 27, 2015
Simulagora – launch the complete study
● In this 1st medium-scale experiment, we had to use a conservative computation power ramp-up (more on this later)
● Real elapsed-time for the 400 computations was 2,5 days
(total of 33 CPU days on the chosen servers - 8Go RAM, 4 CPU)
● We can now achieve the same within ≃ 5 hours (and soon a bit less)
Number of running computations over time
Collaboration:Aster Pronet – March 27, 2015
Phimeca's feedback (1/2)Code Aster:
● 1h30 per simulation● RAM 10 GB consumption● 400 simulations to launch
Simulagora vs. clusters:
● Phimeca's internal resources may be temporarily exhausted like during the Sicodyn project
● Simulagora's pros vs. usual external clusters approach = simplicity
– No configuration or install
– No job manager
– Powerful Python API
For this kind of projects, we usually distribute our calculations on external clusters.
Collaboration:Aster Pronet – March 27, 2015
Phimeca's feedback (2/2)Pros:
● Highly and skilled technical support● Machine image customization to come● Python API to distribute the computations and retrieve results (still in
development but very promising)● Web interface to manage the studies
To improve:● Need an even simpler Python API:
– 'Simulagora run script.py' to create and launch a study– 'Simulagora download' to retrieve results
● No commercial offer yet
Collaboration:Aster Pronet – March 27, 2015
Lessons learned by Logilab
● Logilab faced and solved or worked-around real-world problems
– A bug in disk resource freeing● Warned by Amazon because of a 20To limit reach● Resources first freed manually● Quick code fix
– Some computations crashed due to a high load on the web server● Worked-around by staggering the launches● Crashed runs re-init not easy, but a quick work-around
→ clean fix and API improve to come
Collaboration:Aster Pronet – March 27, 2015
Near future of Simulagora
● The Python client will shortly provide higher level operations(suggestion from Phimeca)
● Logilab will now work on a commercial offerWe now think the service has a commercial value
● Beta-testers are still welcome !
Collaboration:Aster Pronet – March 27, 2015
Thanks for your attention!
Florent Cayré[email protected]
+33 1 45 32 03 12www.logilab.fr
Karina [email protected]
Felipe Aguirre [email protected]
+33 1 58 51 18 11www.phimeca.com