Home About Research Seminars Publications
banner

Upcoming Talks

Evaluating Performance with SPEC Standard Benchmark Suites
Speaker:Dr. Junjie Li, Research Technology, Indiana University
When: 3:00 pm - 4:00 pm Thursday (October 12, 2017)
Where: Room 161A, Smith Research Center or join online using ZOOM (https://IU.zoom.us/j/679375838)

Abstract

The Standard Performance Evaluation Corporation (SPEC) is an organization for creating industry-standard benchmarks to measure various aspects of computer performance. SPEC High Performance Group (HPG) is a working group dedicated to develop application benchmark suites for high performance computers, these suites are capable of gauging performance of multiple parallel and accelerator programming models including MPI, OpenMP, OpenACC and OpenCL with optional energy measurement. The latest update to Accel benchmark suite has also added support for OpenMP 4.5 with the capability of target offload. In this talk, I will outline features of SPEC benchmark suites and run rules, discuss the latest benchmark results on various platforms including Xeon Phi and GPU using both OpenMP 4.5 and OpenACC, and how Research Technology, potentially other researchers at IU, benefits from these benchmarks.

Biography

Judy Qiu

Junjie Li is a principal system analyst in Research Technology at Indiana University. He received his PhD degree in chemical physics from the Indiana University Bloomington (2016). During his graduate research, he was key contributor to multiple computational chemistry methods in Gaussian quantum chemistry package including quantum nuclear dynamics and fragment-based ab initio molecular dynamics. At Research Technology, Junjie primarily focus on performance tuning and user support on IU’s high performance computing platforms. He also serves as secretary of High Performance Group (HPG) in Standard Performance Evaluation Corporation (SPEC).

Previous Talks

Machine Learning Architectures for HPC
Speaker: Prof. Minje Kim, ISE, School of Informatics, Computing and Engineering
When: 3:00 pm - 4:00 pm Thursday (September 14, 2017)
Where: Room 161A, Smith Research Center or join online using ZOOM (https://IU.zoom.us/j/679375838)

Abstract

Training machine learning models from big data using iterative training algorithms is one of the most computationally heavy tasks in cloud computing these days. For example, the job can be done involving massive (dense) matrix computations with the support from GPU computing. In this talk I introduce some machine learning models that are more suitable for best utilizing HPC resources (such as KNL). First, I introduce Mixture of Local Experts (MLE), or Modular Neural Networks (MNN), which combine heterogenous specialized networks as modules. We look into this kind of models and see its potential advantages in parallel computing. Second, we also investigate the sample distribution of the network parameters, and their use in scheduling algorithms for neural network training. An immediate idea would be to update more important parameters more frequently to achieve faster convergence. Finally, I also introduce a binarization technique for probabilistic topic modeling so that the iterative training algorithm can be replaced with a non-iterative bitwise operations, which could be a big save of energy and time during training.

Biography

Judy Qiu

Minje Kim is an assistant professor in the Dept. of Intelligent Systems Engineering at Indiana University. He received his PhD degree in Computer Science from the University of Illinois at Urbana-Champaign (2016). Before joining UIUC, he worked as a researcher in ETRI (a national lab in Korea) from 2006 to 2011. He did his Bachelor’s and Master’s studies in the Division of Information and Computer Engineering at Ajou University (honor) and in the Department of Computer Science and Engineering at POSTECH (summa cum laude) in 2004 and 2006, respectively. His research focuses on designing machine learning models and applying them to signal processing problems, stressing out their computational efficiency in the resource-constrained environments or in the implementations involving large unorganized datasets. He received Richard T. Cheng Endowed Fellowship from UIUC in 2011. Google and Starkey grants also honored his ICASSP papers as the outstanding student papers in 2013 and 2014, respectively.

Affiliated sites Contact
Thomas Wiggins
email: wigginst(at)indiana.edu