CS Seminar: Performance Modeling and Prediction
  • FENS
  • CS Seminar: Performance Modeling and Prediction

You are here

Speaker: Murat Tıkır (Google Inc.)

Title: Performance Modeling and Prediction

Date/Time: April 19, 13:30--14:30

Place: FENS L035

 

Abstract: Performance models explain the performance characteristics of applications running on underlying architectures. The most common application for a performance model is to predict the runtimes of applications when the input data set changes and/or the application scales to larger task counts. This makes models to be frequently employed by computer vendors in their design of future systems by constructing a model for one or two key applications, and then comparing future technology options based on model projections. Performance models have also been effectively used in system and application tuning, making informed decisions on benefits of porting applications to new systems/hardware, and finally in procurement process of selecting the optimal system for a given set of workload in national and university labs.

 

In this talk, I present my prior research in performance modeling and prediction based on PMaC Prediction Framework, an implementation of an automated prediction model. In this framework, a model is a calculable expression that takes attributes of application software, input data, and target machine hardware as parameters and computes the expected performance as output. I will first present a scheme for predicting the performance of HPC applications based on the results of simple benchmarks measuring the performance of underlying memory subsystem. In this scheme, a Genetic Algorithm approach is used to learn bandwidth as a function of cache hit ratios per machine using the simple memory benchmark as the fitness test. Then, I will discuss PSINS, an efficient, accurate and flexible trace-driven performance modeling and prediction tool for MPI applications. A principal feature of PSINS is its usability for applications that scale up to large processor counts while producing accurate predictions. This provides a flexible framework for collaboratively exploring the implications of constantly growing supercomputers on application scaling, in the context of network architectures and topologies of state-of-the-art and future planned large-scale systems. I will also demonstrate that the combination of these tools can effectively be used to predict application performance in existence of hardware accelerators and be used to model and predict power footprints similar to the runtimes. I will conclude my talk with a case study scaling an earthquake simulation application to almost 150K cores based on iterations of modeling and prediction repeatedly, which is nominated as a finalist for Gordon Bell Prize.

 

Bio: Mustafa Murat Tikir received his PhD degree at the University of Maryland, College Park in September 2005. He received his BS degree at the Middle East Technical University, Ankara, and MS degree at the University of Maryland, College Park. His research interests are in the areas of High Performance Computing, Programming Languages, Systems, Web & Mobile Services, Cloud Computing/Big Data. He is interested in performance prediction, performance modeling and tuning of large scale applications with large input sets. His PhD research developed several profile-driven techniques to dynamically increase the locality of memory accesses in memory-intensive applications running on cc-NUMA architectures. Following PhD, he has joined Performance Modeling and Characterization Lab (PMaC) at Supercomputer Center in the University of California, San Diego, as a researcher bringing the scientific rigor to the prediction and understanding of factors affecting the performance and power footprints of HPC platforms. He has primarily worked on performance modeling and prediction of large scale scientific applications. Later, to also obtain industry experience, he has joined Google, Inc in November 2010. He has worked in Google Analytics measuring web site user perceived latencies and Personal Search where users can search their private data from Google properties. He currently is working as a Staff Software Engineer contributing to Google Now, a reliable personal assistant with rich personal predictive content based on user needs.