Elastic High Performance Applications – A Composition Framework
0 likes385 views
The document presents a framework for elastic high-performance applications (EHPA) aimed at addressing the complexities of software dependencies and elastic requirements in cloud environments. It highlights the importance of multi-dimensional elasticity, including resources, quality, cost, available time, and rights of use, and outlines composition algorithms to resolve dependencies and form partitions. Future work emphasizes the integration of modeling techniques and the need for further development in runtime packaging and deployment strategies.
Elastic High Performance Applications – A Composition Framework
- 1. Elastic High Performance Applications – a Composition Framework Tran Vu Pham1, Hong-Linh Truong2, Schahram Dustdar2 1 Faculty of Computer Science and Engineering Ho Chi Minh City University of Technology 2 Distributed Systems Group, Vienna University of Technology truong@infosys.tuwien.ac.at http://www.infosys.tuwien.ac.at/Staff/truong APSCC 2011, 13 Dec 2011, Jeju, Korean 1
- 2. Outline Motivation Elastic components and elastic high performance applications Prototypes and experiments Conclusions and future work APSCC 2011, 13 Dec 2011, Jeju, Korean 2
- 3. Background APSCC 2011, 13 Dec 2011, Jeju, Korean 3
- 4. Motivation (1) HPC cloud markets: Infrastructure Providers, Software Providers, Service Vendors, and End-user Diverse types of components: HPC programs, libraries, operating systems, virtual machine images, Web services, SaaS, PaaS, and IaaS, Different costs and licensing Complex application requirements: quality, elastic time and money, scale in/out different cloud environments APSCC 2011, 13 Dec 2011, Jeju, Korean 4
- 5. Motivation (2) Existing resolving software dependency and compatibility → deal wih software dependencies and compatibilities around a fixed OS Workflow composition → deal with matching service input/output eHPA: → conflicting diverse types of components within and among cloud-based environments dynamically Few existing solutions deal with resource elasticity only → there are multi-dimensional elasticity (see “Principles of Elastic Processes – IEEE Internet Computing 15, 5 (September 2011), 66-71”) APSCC 2011, 13 Dec 2011, Jeju, Korean 5
- 6. eHPA components and relationships APSCC 2011, 13 Dec 2011, Jeju, Korean 6
- 7. Multiple elasticity dimensions for eHPA Resource elasticity: software (os/library/middleware/servic Non-functional parameters e) on multiple clouds elasticity: quality, available time, right of uses Pricing/Rewarding/Incentive elasticity: cost eHPA elasticity See multiple elasticity dimensions at: http://www.slideshare.net/linhsolar/principles-of-elastic- processes-on-clouds-and-some-enabling-techniques These elasticity metrics are simplified for the sake of brevity APSCC 2011, 13 Dec 2011, Jeju, Korean 7
- 8. Requirements and elastic properties Functional requirements Properties End user/Service Vendor Software/Infrastructure Provider Functions + + Dependencies + Conflicts + Elastic properties Properties End user/Service Vendor Software/Infrastructure Provider Resource + Cost + + Quality + + Time + + Rights of Use + + APSCC 2011, 13 Dec 2011, Jeju, Korean 8
- 9. Elastic Component and eHPA Elastic component Functional description Elastic properties Elastic High Performance Application (eHPA) Component properties and dependencies are modeled using ontology APSCC 2011, 13 Dec 2011, Jeju, Korean 9
- 10. Elastic measurements and aggregation Resource for components Internal dependency External dependency Component cost Aggregated cost Quality Available Time Right of Uses APSCC 2011, 13 Dec 2011, Jeju, Korean 10
- 11. Composition algorithms (1) Requested partitions of components and elastic requirements eHPA Compostion – functionality aspect Resolve dependencies, check conflicts, and form partitions EHPA composition – elastic requirement Check elastic requirements APSCC 2011, 13 Dec 2011, Jeju, Korean 11
- 12. Composition Algorithms (2) APSCC 2011, 13 Dec 2011, Jeju, Korean 12
- 13. Prototype APSCC 2011, 13 Dec 2011, Jeju, Korean 13
- 14. Experiments – application Star3D Solving Euler equations in the cases of 3D flows Based on MPI APSCC 2011, 13 Dec 2011, Jeju, Korean 14
- 15. Experiments – modeling Star3D Elastic properties Resources: 32 processes Subjective rank: 2-4 Free of charge Unlimited time Academic license APSCC 2011, 13 Dec 2011, Jeju, Korean 15
- 16. Experiments – possible solutions 75 components in knowledge based Star3D on EC2 with linux 12 different solutions Four groups of solutions, different component external and internal dependencies APSCC 2011, 13 Dec 2011, Jeju, Korean 16
- 17. Experiments – examples of solutions An eHPA solution using free a Fortran compiler (solution 8) An eHPA solution using Portland Fortran compiler, licensed for use up to 256 MPI processes (solution 4). APSCC 2011, 13 Dec 2011, Jeju, Korean 17
- 18. Discussion and future work Complex HPAs in clouds Deal with complex software dependencies and conflicts Determine and characterize elastic properties Multi-dimensional elasticity metrics: resource, quality, cost, available time and right of uses We propose modeling and composition techniques We use simple elastic properties but they can further modeled into sub-dimensions our first step toward multi-dimensional elasticity for HPAs Current we do not consider dependencies among these properties Our future work Integrate with TOSCA (www.open-tosca.org) Work on elasticity tradeoff Develop runtime packaging and deployment APSCC 2011, 13 Dec 2011, Jeju, Korean 18
- 19. Thanks for your attention! Hong-Linh Truong Distributed Systems Group Vienna University of Technology Austria truong@infosys.tuwien.ac.at http://www.infosys.tuwien.ac.at/staff/truong APSCC 2011, 13 Dec 2011, Jeju, Korean 19