Time granted. -- John Roberts Argonne National Laboratory CELS Systems [email protected] On 2/20/17, 3:00 PM, "[email protected] on behalf of [email protected]" <[email protected] on behalf of [email protected]> wrote: Hello, A change in allocation has been requested: Requester: pbalapra (Prasanna Balaprakash) Project: perfopt Title: Optimization Algorithms for Empirical Performance Description: There are different strategies for performance model construction. Analytical approaches for performance modeling are cheaper in the sense that they don't suffer from the profiling and model generation cost as in the case of machine learning techniques. But analytical models are often difficult to construct for complex programs and they require detailed understanding of the underlying hardware and implementation of communication functions. Machine learning techniques come into play when it is difficult to construct an analytical model for an application. Models generated using machine learning techniques may not be as highly precise as an analytical model but still they help to understand the performance of an application that in turn help in optimizations such as auto tuning. In this research, we first try to build analytical models for communication of various application kernels. We use architecture-independent parameters as described in well known communication models such as LogGP family to build the analytical models. If the constructed analytical model is accurate enough in predicting the communication time, it saves the huge overhead in gathering training data for the application of machine learning algorithms. We will test the accuracy of the developed analytical models in predicting the communication time of several application kernels from the CORAL benchmarks. If the analytical model is not accurate enough (the error in prediction goes above acceptable range), we switch to gathering profile data and then generate models using machine learning techniques on that data. This hybrid approach gives us a balance between the complexity of building analytical models and the overhead of running expensive machine learning algorithms on huge amount of collected profiling data. Also the machine learning models are limited in the sense that we can't do extrapolation. Fitted analytical models come into play in that case. On Blues, we will conduct an experimental study of communication models. A major benefit Blues brings to this research is the ability to test the impact of MPI implementations and interconnect techniques on Intel-based clusters. Eventually, we hope to deploy our techniques for improving the performance of scientific applications running on Blues. We estimate 100,000 core hours for our experimental study: evaluation of the strengths and limitations of LogGP models = 15,000 core hours; experimental analysis on multiple problem classes on different architectures = 30,000 core hours; empirical analysis for new approaches = 30,000 core hours; tests on parallel performance tuning approaches = 25,000 core hours; We will strongly involve in research dissemination activities such as writing scientific articles and giving seminars at Argonne and in conferences. In the the first and second quarter, our jobs will be 80% parallel. For the third and fourth quarter, we will have 100%. Current: undetermined amount Justification: Requested: 80000 A specific reason has been given: Q2: 15000 Q3: 40000 Q4: 25000 This needs to be approved and the final allocation amount decided upon. Thank You, The LCRC Accounts System _______________________________________________ allocations-admins mailing list [email protected] https://lists.lcrc.anl.gov/mailman/listinfo/allocations-admins