Charlene Wang

SC17 Conference Summary

Blog Post created by Charlene Wang on Nov 17, 2017

SC17 Conference Summary

Event Detail: Nov 12-17 2017 Denver, Colorado

Attendee:

affliation

member

SE

Hiroshi Utsunomiya(Public Platform Solution)

Sales

Naohiro Sekiguchi

Sales

Shuichi Minowa

SE

Kyoichi Saito

SE

Naofumi Ishikawa

Lab

Masaaki Shimizu

SE

Yuji Kojima

Sales

Kohe Hashimoto(Booth Master)

HPC products developer

Tadayuki Sakakibara

HPC products developer

Yasuhiro Teramoto

SE

Atsuhiro Suzuki(Lumada)

HPC products developer

Mitsuyoshi Ishihara

Sales

Hirotaka Niikura

Hudson Engineer

Shinya Imaizumi

Hudson Engineer

Tsukasa Hosoya

Lab

Kazuhisa Fujimoto(HPC Architect)

Lab

Toshiyuki Aritsuka

HAL

Yuki Sakashita(Deep Learning)

Vantala

Satomi Hasegawa(Lumada Product Management)

Vantala

Charlene Wang

Vantala

Erik Urdang

 

Observation Summary:

The main theme this year is AI/ Machine Learning, featuring latest advancement in HPC infrastructure, machine learning software and software-driven cloud computing.

 

HPC infrastructure highlights:

  Vendors feature Nvidia/ Intel/ AMD GPU, various FPGA accelerator and Rasberry Pi supercomputer, IoT appliance. Some vendors talk quantum computing.

  Hitachi showcased

  1. 1. Deep Learning Server with max 6 Nvidia V100 GPU and IBM POWER9 and a Fortran compiler that auto port existing Fortran programs from CPU to this GPU server.
  2. 2. (Intel) FPGA accelerator (hardware, middleware and plugin) for Hadoop SQL engine results in 100 times faster query response time. 
  3. 3. Machine learning software:

  Most vendors showcase performance benchmark with open source Tensorflow, Caffe, Spark. Mathworks supports deep learning with proprietary software.

  e.g.

         NEC :

                  demo vector engine -> Spark -> recommender for movies

         Fujisu :

                  full immersion warm water cooling

                  DLU ( not GPU ) that supports tensorflow, caffe

         BitScope :

Demo a super conmputer with 200 Rasberry phi. cheapter and lower power consumption makes it ideal testbed for student lab.

         IBM :

                  Instead of demo Watson, IBM demo distributed deep learning application management tool for researchers to collaborate development using Python Jupyter, tensorflow, caffe1 ( does not support caffe2 yet ).

         DellEMC :

Showcased a generative adversarial network (GAN) that transfer real-time camera video into Renaissance Artists’ art style and display. The software was developed by Nvidia a year ago.

         Google cloud :

                  Different scientists, researchers share their projects, user experience.

         D-wave : a quantum computing co.

Tech Talk Highlights:

  1. 1. Keynote speakers presented SKA - a collaboration project from government around the world to understand universe, dark material, gravitational wave, etc.

One thing I learned from the talk is in SKA data pre-processing: they synchronize signals from all satellite channels before cleaning and training model.

  1. 2. New business model: HPC infrastructure as a service and Analytics as a service.
  2. 3. Machine learning proof of value metric can include trend, scalability in addition to accuracy.
  3. 4. Active learning may be able to simplify experiments design.
  4. 5. Use precision/ recall to choose threshold and model.
  5. 6. Performance modeling under resource constraints using deep transfer learning.
  6. 7. Characterizing of faults, errors, and failures in extreme-scale systems

(Hideyuki Jisumoto Tokyo Tech, Lawrence Livermore Lab, Oak Ridge Lab, Swiss National Supercomputing Center).

         -resilience is a cost optimization problem.

         -distinguish transient failure

         -degree of failure

         -guaranteed by statistical method

         -identify cause contribute to variability

         -need manual interactive annotation for operation logs to label faults, errors, and failures.

  1. 8. China Sunway Taihulight is world's largest HPC with 10 million core.

   one use case featured by Sunway speaker: Sunway taihulight enabling depiction of realistic 10hz scenario, 15-Pflops nonlinear earthquake simulation. Techniques used to increase memory bandwidth include:

                               multi-level domain decomposition

                               DMA and buffering scheme

                               seismic data compression: squeezing extra performance

   Sunway Taihulight project was funded by Chinese government that include design and develop chip, server, and software.

         Deep Learning on Sunway Taihulight: currently supports swDNN (vs Nvidia cuDNN) and swCaffe with plan to port Baidu's paddle paddle.

   Currently no application running on this supercomputer yet.

   Government cut funding so they need to convert some research to commercially available product to fund next phase development.

Outcomes