Login
Talks & Seminars
Title: Systematically and Algorithmically Tuning Massively Parallel Applications
Dr. Preeti Malakar, Argonne National Labs
Date & Time: July 27, 2017 14:30
Venue: Conference Room, C Block, 01st Floor, Dept. of CSE, Kanwal Rekhi (KReSIT) Bldg.
Abstract:
Computational science helps in solving fundamental problems in various scientific fields such as climate and weather, material science, drug discovery and many others. The complexity of the mathematics and the large number of calculations involved in modeling these real-world scientific problems necessitate the use of high performance computing. This has led to the development of current petascale supercomputers that have enabled the execution of transformative scientific simulations at unprecedented scales. Critical applications such as cyclone tracking require accurate and timely predictions. It is also necessary to identify important regions of interest and simulate them in greater detail. Simulating dynamic regions of interest entails massive amount of inter and intra-node communications. The first part of this talk will focus on communication optimizations in a weather application by improving processor allocation and reallocation strategies for nested simulations which also improves the performance of the simulation. I will describe a strategy for parallel execution of multiple nested domain simulations based on partitioning the 2D process grid into disjoint rectangular regions associated with each domain. The strategy consists of a novel combination of performance prediction, processor allocation and reallocation methods, and topology-aware mapping of the regions on torus interconnects. Experiments show 33% performance improvement over the default strategy. Next, I will focus on communication optimizations in the context of parallel I/O from thousands of cores. Optimizing data movement in exascale systems is challenging due to complex network topologies, proprietary routing algorithms, and dynamic application communication patterns. I will describe my work on topology-aware data movement for I/O at scale. Scientists can achieve faster time-to-solution for their science by reducing the I/O time. I will outline our route-aware and load-aware algorithm to modify the existing I/O node assignments in the Blue Gene/Q supercomputer during runtime. This mitigates network contention and reduces write times by an average of 60% over the default MPI independent I/O on up to 8192 Blue Gene/Q nodes. The algorithm routes 1.4x fewer messages through the I/O nodes on the Blue Gene/Q. I will conclude with a brief description of our work on scalable visualization to help identify bottlenecks in massively parallel communications
Speaker Profile:
Preeti Malakar is an Assistant Computer Scientist in the Argonne National Laboratory, after spending three years as a postdoc at Argonne. Her current work includes topology-aware performance optimizations for in situ analysis of scientific simulations and large-scale parallel I/O on some of the fastest supercomputers. She received her PhD from the Indian Institute of Science, Bangalore in 2014. She was awarded the Google India Women in Engineering Award in 2011 and the TCS Research Fellowship during her PhD.
List of Talks

Webmail

Username:
Password:
Faculty CSE IT
Forgot Password
    [+] Sitemap     Feedback