GPUs and Parallelization - Biye Jiang, Aaron Culich

October 21, 2015 at 5-6:30pm in BIDS, 190 Doe Library


Meeting Info

Biye Jiang

Biye Jiang is a PhD student at UC Berkeley in the CS department working with John Canny.

Aaron Culich

Aaron is a research computing architect at Berkeley.

Discussion: GPUs and Parallelization

Today’s topic is about GPUs and parallelism.

Survey of Needs and Resources – Aaron Culich

Aaron referenced a presentation on this topic. It can be found here.

Aaron started this presentation with a survey of what the attendees are actually using.

Python Parallelism

It was mentioned that, for some folks, python is the language of choice. The Python Multiprocessing module was mentioned. This was the topic of a THW session last year. The THW resources on this topic can be found here. That session was not on GPUs, however, the python threading module can be used in conjuction with PyCUDA, a python module for GPUs.

Research IT – Krishna Muriki

Research IT is available as a resource for individuals who would like to test their code on GPU resources. Krishna Muriki expresses that there is an institutional shared linux cluster (Savio). Within that cluster, there are 6 compute nodes with 4 kepler GPUs each. Those nodes are in testing and BRC is interested and open to new users.

Java runtime engine – Oliver

Oliver at ESPM has a javascript modeling project for agent based population models. They are working to make their software scalable from the desktop to the level of higher performance computing. The NOVA stack and XSEDE resources are core to their efforts.

Scala Demo – Biye Jiang

Biye demonstrated the speed of GPUs by conducting a matrix multiplication using GPUs versus conducting the same multiplication using CPUs.

GPU Discussion

Biye shared some of the diagrams from this presentation.

He noted

GPU BIDMat demo

Biye presented an ipython notebook to demostrat how BIDMat works.

The ipython notebook demos are here.

Lightning Talks

Finally, there will be a time for a couple of Lightning Talks, which are 5-10 minute blasts of information about a particular topic or question of interest to the group. This topic can be anything useful, new, or interesting to scientists who compute. It may be some new skill you have recently picked up in your research, a productivity tool you have recently learned to love, a quick demo of a useful library, or anything you feel we would enjoy learning.
Note that the lightning talk time is a good way to bring a question to the group. If you have a bug you need help with, here’s the place to ask many ears about it at once.

Name : Topic

Notes and links

Name : Topic

Notes and links