Runtime Autonomous Component Management Systems Runtime Autonomous Component

Runtime Autonomous Component Management Systems

Runtime Autonomous Component Management Systems n n CMS Runtime Component Optimizer n We have designed software APIs for CMS Runtime Optimizer n Develop general-purpose and application specific Runtime Optimizer using these APIs Switching Supports for CMS n Load balancing Mechanism n Supports Stateful and Stateless invocation n Work-flow model

Runtime Autonomous Component Management Systems n SIP/Mobile RMI integration n Protecting transmitted data when composing components dynamically in a distributed environment n Supporting more types of mobility using Session Initiation Protocol (SIP) and Mobile RMI complementally n Combine both mechanism to a single integrated mechanism to reduce overhead and redundancy n Integrate SIP servers with Mobile RMI facilities

Mechanism Diagram Client 1 Client 2 Cluster 1 Linux Host CMS Cluster 2 Network Processor CMS Client 1 Client 2 CMS CMS Cluster 1 Cluster 2

. NET Remoting n n n An emerging object-oriented remote-procedure-call protocol under Microsoft’s. NET framework, like RMI in JAVA Let programmer concentrate on business logic instead of socket operation in network application Three Service Activated Mode: Both Singleton and Client. Activated modes provide the stateful access Client Side Server Side Client Side C 1 S 1 C 2 Server Side S 1 Client Side Server Side C 1 S 1 C 2 S 1 Single Call Singleton Client-Activated

Load Balancing Mechanisms n Start In TCP connection table yes no n Find service type Stateful? no yes Dispatch We propose two methods: Find a In Session no least load table? server yes Insert Apply session previous server table n ETT (Estimated-Task -Time) Scheduling Method EFT (Earlier-Finished -Time) Scheduling Method : Consider the Workflow Graph

ETT Scheduling Method n n n We use the Estimated-Task-Time model to calculate the cluster states at run-time. Use given computation cost and communication cost to estimate the total execution time to be consumed. Calculating the ETT values of each cluster to find out the server for assignments.

EFT Scheduling Method --Handle Workflow Graph n n To cope with application when given a workflow graph, we propose a scheduling policy. We use a two-phase algorithm: n n n Phase 1: Schedule the stateful tasks. Phase 2: Schedule the remained stateless tasks and dispatch the stateful tasks which have been scheduled at phase 1. Use a Time-Out policy to re-schedule stateful tasks when timeout constraints are met during phase 2.

Phase 1 of EFT algorithm n n We balance the total load of each stateful task groups to each server. Decide the target server for tasks to be assigned.

Phase 2 of EFT algorithm n n n We use the rank function to find out the critical path to decide the execution order. Use Earlier-Finish-Time Function to find out the server to finish the task early. If the stateful group has met time-out constraint, the group of the tasks will be rescheduled. Assign task by the Rank order Stateful no Schedule task by EFT function no Schedule task by phase 1 decision yes Time out yes Restart Phase 1 Dispatch

Experimental Result n n We examine our implementation in IXP 1200 which was compared with Microsoft NLB. We experiment with our workload algorithms EFT by simulations. We use 500 graph instances for evaluating each parameter settings. The performance results with ETT and EFT is normalize to the results of Round-robin (RR). Parameter Value V 25, 50, 100, 200, 400 S 1 O 2, 3, 4 CCR 0. 3 Stateful groups 2, 4, 8 Stateful task ratio 0. 25, 0. 5

Experimental Result Experiment result of 25% stateful task graph sets 25 50 100 200 400

Experimental Result Experiment result of 50% stateful task graph sets 25 50 100 200 400

Summary n n n We proposed a load balancing methodology which supports stateful service access. We can see that the EFT algorithm has significant performance improvement compared to ETT and RR. While the stateful task ratio is 50%, the improvement of EFT is from 5% to 21% when compared to ETT and is from 8% to 34% when compared to RR.

本年度計畫產出物 論文 – n n n “Efficient Switching Supports of Distributed. NET Remoting with Network Processors. ” C. K. Chen, Y. H. Chang, C. W. Chen, Y. T. Chen, C. C. Yang, and Jenq-Kuen Lee. ICPP 2005. “Switching Supports for Stateful Object Remoting on Network Processors”, C. K. Chen, Y. H. Chang, Y. T. Chen, C. C. Yang, and Jenq Kuen Lee, accepted, Journal of Supercomputing, (Special Issue for Selected Papers of CTHPC 2005). „Mobile Java RMI Support over Heterogeneous Wireless Networks“, C. K. Chen, C. W. Chen, C. T. Ko, Jenq-Kuen Lee and Jyh-Cheng Chen, Submitted to IEEE Transcation on Mobile Computing. 專利 1. 1. 2. Jenq-Kuen Lee, Jyh-Cheng Chen, Cheng-Wei Chen, Chung-Kai Chen, “Method and System for Providing Roaming of Remote Object Procedure Call in Heterogeneous Wireless Network Environment, Pending Patent (已申請台灣專利, 美國專利申請中). “Mechanism for Supporting Stateful Object Remoting”, 準備提 出專利申請 (台灣及美國).