The OSCAR framework is previewed in order to automatically deploy an elastic Kubernetes cluster with Minio (as the storage back-end), OpenFaaS (as the FaaS framework to execute functions in response to events), Event Gateway (to route events). Open-source development in GitHub: https://github.com/grycap/oscar

1. ON-PREMISES SERVERLESS
CONTAINER-AWAREARCHITECTURE
(OSCAR)
Alfonso Pérez, Sebas Risco, Germán Moltó,
Miguel Caballer,AmandaCalatrava
IBERGRID 2018
October 11-12, Lisbon, Portugal
Instituto de Instrumentación para
Imagen Molecular
Universitat Politècnica de València
Spain

2. GOALS OFTHISTALK
• Briefly describe the open-source developments of the
GRyCAP to support scientific computing on multi-
Clouds and serverless platforms
• From Infrastructure as a Service (IaaS) to Functions as a
Service (FaaS) for Scientific Computing
• Introduce the OSCAR framework to support serverless
computing (FaaS) for file-processing scientific
applications on elastic Kubernetes platforms

3. FROM IAAS TO FAAS
3

4. • FaaS- Functions as a Service (ServerlessComputing)
– Executes functions in response to events with automated elasticity
• Stateless functions executed in containers
• Up to 5 minutes of execution time
• Allows different programming languages:
– Node.JS, Python, Java, C# and Go
• Pros
– No infrastructure provision or configuration management
– Large elasticity (thousands of parallel executions)
• Cons
– Requires re-architecting the application
– Difficult to customize execution environment
PUBLIC FAAS: AWS LAMBDA
4

5. EXTENDINGAWS LAMBDA:
SCAR
• Serverless Container-awareArchitectures
– Runs on AWS Lambda
– Executes customized runtime environments
provided by Docker containers (using uDocker)
– Highly-parallel event-driven file-processing
serverless applications
– Bring your own application and runtime

6. SCAR: ARCHITECTURE
• Supervisor in charge of
retrieving the Docker
image from Docker Hub
• Caching depending on
the Freeze/Thaw cycle of
underlying containers
• Useful for bursts of short
stateless computational
jobs
• Limitations
– Small Docker Images
(⪅ 230 MB)
6
Lambda
S3 CloudWatch
SCAR (Client)
Lambda function
SCAR (Supervisor)
User-defined
Docker image in
DockerHub
Received
event
Lambda function instances
Log
Streams
S3
bucket
/input
/output
Creates
Outputs Logs
initrun
return2 1 3
/tmp
Docker
Hub
pull
Udocker
A. Pérez, G. Moltó, M. Caballer, and A. Calatrava,“Serverless computing for container-based
architectures,”Futur. Gener. Comput. Syst., vol. 83, pp. 50–59, Jun. 2018.

7. USE CASES OF SCAR
• Massive image recognition using Deep Learning /
Neural Network on datasets of images
– Service triggered on-demand upon uploading a file
into a bucket
• Zero infrastructure cost if not used (uses AWS Lambda)
• Supporting languages / applications not being
easily supported by AWS Lambda
– R, Erlang, Elixir
• Image /Video transformation

8. AN OVERVIEW OFTHE SERVICES
http://github.com/grycap
IM
• Automatic deployment of
virtual infrastructures on
multi-Clouds
CLUES
• An elasticity management
system for Clusters and on-
premises Clouds / elasticity
manager
• Elastic virtual clusters
on public and on-
premises IaaS Clouds
EC3 SCAR • Serverless computing
for scientific workloads
OSCAR
• On-Premises Serverless
computing (FaaS) for
Scientific Computing

9. ON-PREMISES SCAR (OSCAR):
COMPONENTS
EC3IM
CLUES
EVENT
GATEWAY
KUBERNETES
Deployment on multi-
Clouds
Elastic Container
Orchestration Platform
MINIO
OSCAR
Services

10. OSCAR ARCHITECTURE

11. DEMOTIME
• Deploy a Plants Classification application so that
every time an image is uploaded to a Minio an
event is generated to Event Gateway and used
to asynchronously invoke an OpenFaaS function
to perform the file processing.
• The results will be made available in another
Minio bucket.

12. DEMOTIME

13. DEMOTIME

14. DEMOTIME

15. DEMOTIME

16. DEMOTIME

17. CONCLUSIONS
• Functions as a Service frameworks provide a layer of abstraction
on top of Container Orchestration Platforms
• In OSCAR, the elasticity of function invocations is handled by the
FaaS to automatically trigger the elasticity of the Kubernetes
cluster thanks to CLUES and the IM to provision additional nodes
of the cluster
• HighThroughput Computing Programming Model
• FutureWork
• Support for multiple data storage back-ends (EGI DataHub, dCache, etc.)
• Web-based GUI
• Hybrid (on-premises / public) workload distribution.

18. ACKNOWLEDGEMENTS
• The authors would like to thank the current projects
BigCLOE[1], DEEP Hybrid-DataCloud[2] and the EGI
Strategic and Innovation Fund for the financial support
to continue the aforementioned developments.
[1] "Ministerio de Economía, Industria y Competitividad” for the project “BigCLOE - Computación Big Data y de Altas
Prestaciones sobre Multi- Clouds Elásticos” with reference number TIN2016-79951-R.
[2] European Union’s Horizon 2020 research and innovation programme under grant agreement No 777435.

19. CONTACT
Alfonso Pérez / Germán Moltó
Instituto de Instrumentación para Imagen Molecular
Universitat Politècnica deValència
Camino deVera s/n
46022,Valencia
SPAIN
alpegon3@upv.es
gmolto@dsic.upv.es
http://www.grycap.upv.es/gmolto @gmolto