60 likes | 181 Views
An increasingly important platform for large scientific applications are collections of distributed resources Resources may include workstation clusters MPPs storage, etc. PROBLEM: How can users schedule applications to achieve performance in multi-user, distributed platforms?
E N D
An increasingly important platform for large scientific applications are collections of distributed resources Resources may include workstation clusters MPPs storage, etc. PROBLEM: How can users schedule applications to achieve performance in multi-user, distributed platforms? SOLUTION: Adapt application execution to the resources which offer the best performance T3E Viz Using AppLeS to Improve the Performanceof Distributed Applications cluster data
AppLeS:Adaptive Application Scheduling in Production Distributed Environments • Project Lead: Fran Berman, UCSD • AppLeS = Application-Level Scheduler • Each AppLeS is a scheduling agent for an individual distributed application • Each AppLeS develops and executes a custom application schedule • AppLeS schedules adapt to predicted deliverable resource performance at execution time. • For NAVO PET applications, AppLeS targets applications to Legion platforms • infrastructure based on object-oriented model • AppLeS uses Legion’s Kerberos model for handling security
PET AppLeS Project:Results for Interactive Legion Platform • Demonstration: Developed adaptive AppLeS scheduler for Legion PMHD3D (magneto-hydrodynamics) application • representative regular stencil-based application • AppLeS complements Legion by providing adaptive application scheduling mechanism • AppLeS PMHD3D Scheduler • partitions application based on prediction of deliverable performance of target resources • more work assigned to unloaded processors • Wolski’s Network Weather Service provides resource forecasts • AppLeS dynamically schedules computations on best resources, achieves improved performance over static techniques
Experimental Results on Legion Cluster • Profile of heavily loaded processor, when load is heavy, application performance is likely to be poor • AppLeS scheduler uses Network Weather Service to forecast when load will be heavy, develops custom application schedule which targets tasks to faster, more lightly loaded processors • In experiments on Legion cluster, static scheduler always picks the same target processors no matter what the load • Static scheduler chose this processor whereas AppLeS adaptive scheduler chose faster, more lightly loaded processors CPU Utilization
Improving Application Performance on the Legion Cluster • Comparison of statically-scheduled PMHD3D execution and AppLeS- scheduled PMHD3D execution when multiple users present in system. • AppLeS schedule leverages best resources at execution time, statically scheduled execution cannot • Execution time for statically-scheduled application takes 2.5 times longer on average than adaptive AppLeS-scheduled execution. • Adaptive scheduling key to performance improvement for PMHD3D.
Proposed Year 4-5 Activities • Project Goal: Develop AppLeS schedulers to minimizeturnaround time = wait time + execution time by leveraging both batch and interactive environments • Project Goals: • Develop AppLeS to target T3E and other batch MPPs • Develop strategy for predicting execution time on available batch and interactive environments • Develop AppLeS which reduces application turnaround time by running on batch MP, interactive cluster, orbothenvironments simultaneously -- whichever will deliver the best performance • AppLeS Home page • http://www-cse.ucsd.edu/groups/hpcl/apples/