Presentation is loading. Please wait.

Presentation is loading. Please wait.

IEEE CLOUD’2012 Topology-Aware Deployment of Scientific Applications in Cloud Computing Pei Fan 1, Zhenbang Chen 1, Ji Wang 1, Zibin Zheng 2, Michael R.

Similar presentations


Presentation on theme: "IEEE CLOUD’2012 Topology-Aware Deployment of Scientific Applications in Cloud Computing Pei Fan 1, Zhenbang Chen 1, Ji Wang 1, Zibin Zheng 2, Michael R."— Presentation transcript:

1 IEEE CLOUD’2012 Topology-Aware Deployment of Scientific Applications in Cloud Computing Pei Fan 1, Zhenbang Chen 1, Ji Wang 1, Zibin Zheng 2, Michael R. Lyu 2 1.National University of Defense Technology, China 2.The Chinese University of Hong Kong, Hong Kong, China 1

2 Outline  Background  Topology-aware deployment  Experiments  Conclusion 2

3 Scientific Applications  Scientific Applications a.Mathematical models and numerical techniques to solve scientific, social scientific and engineering problems b.Needs huge computing resources c.Involves a number of distributed components d.Requires high bandwidth for data transportation  Topology-Aware Applications a.Scientific applications are topology aware b.Components are connected according to a specific topology 3

4 Communication Graph 4 0-7 represent the distributed components White arrows represent the message exchanged between components

5 CLOUD Computing  A possible solution for providing a flexible, on- demand computing infrastructure for scientific applications.  Computing and storage resources  Geographically distributed datacenters Internet Datacenter1Datacenter2 5

6 Deployment on Cloud (1)  Deploy software components of scientific applications on selected cloud nodes. 6 Deployment

7 Deployment on Cloud (2)  Traditional deployment approaches a.Random b.Ranking c.Clustering  Ranking is suitable for computing-intensive applications  Clustering is suitable for communication- intensive applications 7

8 Deployment on Cloud (3)  Topology-aware deploy method 8

9 Deployment on Cloud (4)  How to obtain communication topology? 1.Described by designer? 2.Detect communication topology automatically Difficult 9

10 Topology-aware Nodes Selection Framework 1 2 3 10

11 Logical Topology Discovery(1)  Adjacency matrix  The communication topology discovery problem can be modeled as a graph clustering problem A B C D 3 3 1 1 Undirected graph Adjacency matrix 11

12 Logical Topology Discovery(2)  Multi-scale clustering a.Hierarchical clustering algorithm b.Multi-scale refinement algorithm 12

13 Physical Topology discover (1)  Cloud nodes have a special topology P.Fan, J. Wang, Z. Chen, Z. Zheng, M.R. Lyu, “A Spectral Clustering-based optimal Deployment Method for Scientific Applications in Cloud”, Int’l Journal of Web and Grid Services. Vol. 8, No.1. pp31-55, 2012. 13

14 Physical Topology discover (2)  Mapping application processes to nodes  Rank the clusters and the cloud nodes in each cluster  Deploy the scientific application to cloud nodes 14

15 Experiment (1)  Experiment Setup a.experiment environment: PlanetLab b.Use 100 nodes serve as cloud nodes c.Framework ran about 53 days  NPB (NAS Parallel Benchmarks) Benchmark  Widely used MPI benchmark  Evaluate the performance of supercomputers 15

16 Experiment (2)  Performance Comparison a.Use pre-execution and logical topology discovery to obtain the topology structures of the programs in NPB 16 Table 1. The Number of logical topology of NPB programs programs in NPB number of logical topology structures number of required cloud nodes

17 Experiment (3)  Performance Comparison  Makespan: The makespan of a job is defined as the duration between sending out a job and receiving the correct result. 17

18 Experiment (4)  Performance Comparison  Throughput: the total million operations per second rate (Mop/s) over the number of processes. 18

19 Experiment (5)  Load Experiment Topology method: deploys an application based on the communication topology Untopology method: deploy application on 1 cluster. 19

20 Experiment (5)  Load experiment Decreased percent of computing performanceDecreased percent of communication performance 20

21 Conclusion  Conclusion  Pre-execute and multi-scale algorithm to obtain communication topology automatically  Topology-aware deployment method to obtain better performance  Future work  Consider the user experiences  Experiments on scientific applications with larger number of components 21

22 IEEE CLOUD’2012 Topology-Aware Deployment of Scientific Applications in Cloud Computing Pei Fan, Zhenbang Chen, Ji Wang, Zibin Zheng, Michael R. Lyu 22


Download ppt "IEEE CLOUD’2012 Topology-Aware Deployment of Scientific Applications in Cloud Computing Pei Fan 1, Zhenbang Chen 1, Ji Wang 1, Zibin Zheng 2, Michael R."

Similar presentations


Ads by Google