SC15 Supercomputing Conference in Austin, TX

This year’s edition of the Supercomputing Conference - SC15 - drew a record-breaking 12,900+ registered attendees organized by the sigHPC of ACM and the IEEE CS (#SC15) to Austin, TX. The technical program spanned six days. Even though the bulk of the conference is still dominated by traditional (small data) HPC topics, newer “Big Data” themes are making inroads. Other highlights included (importance of GPUs for) Deep Learning, and the many workshops and tutorials. At the hardware front, I liked the presentation of the 64-bit ARM chips (in a working server system), continuation of the D-Wave Quantum computing success story, lots of 3D printing applications and demos. Let’s now walk through the conference in chronological order.

The conference organized a flawless bus service to the many hotels housing the close to 13,000 attendees. Below, you see the registration area. All pictures were taken by yours truly (@DirkVandenPoel) and carry my copyright.

Friday 20 November 2015

This completes my personal summary of just the first five days of the SC15 conference. Stay tuned for more updates.

Lots of concurrent session (see picture below):

I decided to attend the “Data-Intensive Applications on HPC Using Hadoop, Spark and RADICAL-Cybertools” Tutorial by

next >

< previous

After lunch, I decided to attend the “MLHPC2015: Machine Learning in HPC Environments” Workshop. The session kicked off with a great keynote presentation by Bryan Catanzaro (@ctnzr, engineer at @BaiduResearch) on “Scaling Deep Learning”. Click here to watch a video by this presenter on the same topic at the NVIDIA booth. Training deep neural networks is very computationally intensive: training one model takes tens of exaflops of work, and so HPC techniques are key to creating these models.

Next, I attended a great talk by Janis Keuper and Franz-Josef Pfreundt (@Fraunhofer) on asynchronous stochastic gradient descent for #DeepLearning titled “Asynchronous Parallel Stochastic Gradient Descent - A Numeric Core for Scalable Distributed Machine Learning Algorithms.” The authors present a novel parallel updating algorithm for SGD, which utilizes the asynchronous single-sided communication paradigm. Compared to existing methods, Asynchronous Parallel Stochastic Gradient Descent (ASGD) provides faster convergence, at linear scalability and stable accuracy.

Next, Steven Young (see picture below) presented the talk “Optimizing Deep Learning Hyper-Parameters Through an Evolutionary Algorithm” (co-authors: Derek Rose, Thomas Karnowski, Seung-Hwan Lim and Robert Patton). Model selection (as architecture construction) through hyper-parameter choices remains a tedious and highly intuition driven task. To address this, Multi-node Evolutionary Neural Networks for Deep Learning (MENNDL) is proposed as a method for automating network selection on computational clusters through hyper-parameter optimization performed via genetic algorithms.

Next, the talk “LBANN: Livermore Big Artificial Neural Network HPC Toolkit” was presented (authors: Brian Van Essen, Hyojin Kim, Roger Pearce, Kofi Boakye and Barry Chen).

The next talk in the MLHPC workshop was titled “HPDBSCAN – Highly Parallel DBSCAN” by Markus Götz, Christian Bodenstein and Morris Riedel. In this talk, they present a new parallel approach which they call HPDBSCAN.

Next morning, yours truly first attended the SC Workshop on Big Data Analytics (BDAC). The three organizers are shown below: Ranga Raju Vatsavai (North Carolina State University, USA), Scott Klasky (Oak Ridge National Laboratory, USA), Manish Parashar (Rutgers University, USA).

Next, Dr. Lucy Nowell (DoE: US Department of Energy, see picture below) kicked off the event with her talk “Beyond Exascale: Data - The Next Big Thing.” She mainly sees challenges with regard to I/O, more specifically, read speeds of file system. They seem to be stuck at 1 TB/s.

Next, Alexandru Costan (Inria, Rennes, France) presented a talk on using Spark across hybrid clouds.

Building on the same theme as in the morning session, I decided to go to the tutorial on parallel I/O for data-intensive science. This tutorial teaches us about the ADIOS framework.

Next, I decided to walk the huge exhibit hall featuring 343 exhibitors from industry, academia and research organizations from around the world.

The third day started with a plenary session. The general conference chair - Dr. Jackie Kern of the University of Illinois at Urbana-Champaign - welcomed the participants.

Next, the award ceremony was held. The list of distinguished recipients: Dr. Mateo Valero (Universitat Politécnica de Catalonya), Dr. Katherine Yelick (Lawrence Berkeley National Laboratory) and Dr. Alexander Szalay (Johns Hopkins University).

Next, Alan Alda delivered a nice keynote speech on the importance of improved communication of scientific results.

Subsequently, I attended a very interesting technical talk at the D-Wave Systems (@DWaveSys) booth. The Vancouver (Canada)-based company building the first useful quantum computer.

Some more exhibit floor pictures are shown below.

Notice: The head of the person shows up twice in the picture thanks to 3D scanning and printing.

Most of the exhibit-hall sessions were also very well attended (see below).

Next up was the opening plenary talk mainly by Intel researchers including a part by Gordon Moore (we all know because of Moore’s law). Click here (or on the picture) for the entire video on YouTube.

Next, it was time for the award ceremony of the different TOP 500 lists.

The fourth day started off with talks by the awardees.

I particularly liked the keynote talk by Dr. Alexander Szalay (see slides below). He argues that "Science is increasingly driven by #Data." We are seeing the emergence of citizen/internet scientists.

Next, Dr. Kathy Yelick (Lawrence Berkeley National Laborabory, see picture below) delivered her award keynote speech. It is clear that the latter two speakers see the change towards more data-driven science. She emphasized the need for streaming #analytics for #BigData.

Next, I attended the panel titled “Supercomputing and Big Data: From Collision to Convergence.” The focus of this session was to situate the problem/open issues. Unfortunately, they did not really get into how to solving it/them.

For lunch, we went to the NVIDIA food court (right across the street from the Austin Convention Center. Again this year, NVIDIA did a lot to support the HPC community.

Ghent University’s DICT members (Kenneth Hoste - @kehoste - and Ewan Higgs, both in the picture below; @HPCUGent) were also very active during SC15 at different booths and workshops. It’s nice to see the close cooperation with @TACC (in the picture they are at their booth). E.g. Dr. McLay will be visiting Ghent in January 2016.

Over lunch, I headed over to NVIDIA’s GPU educators event. The hotel room was packed with academics eager to learn how NVIDIA is helping them teaching GPU technology.

Next, I attended the plenary talk by Dr. Lynn Orr (Under Secretary for Science and Energy at the U.S. Department of Energy, see picture below). He talked about the Quadrennial Technology Review: An Assessment of Energy Technologies and Research Opportunities.

Next, Dr. Thomas H. Jordan (Director of the Southern California Earthquake Center, see picture below) took the stage with a talk on the “Societal Impact of Earthquake Simulations at Extreme Scale.”