OpenCUDA+MPI A Framework for Heterogeneous GP-GPU Cluster Computing
Kenny Ballou
March 20, 2013
Ballou OpenCUDA+MPI
1 GP-GPU Computing Introduction 2 Problems and Solutions 3 A Little About Methodology 4 What’s Been Done
Ballou OpenCUDA+MPI
1 GP-GPU Computing Introduction
Parallel versus Distributed Computing Applications of Supercomputing Who Uses Supercomputing? 2 Problems and Solutions 3 A Little About Methodology 4 What’s Been Done
Ballou OpenCUDA+MPI
Introduction Parallel and Distributed Computing
What is GP-GPU Distributed Computing? Parallel: Processing concurrently
Distributed: Processing over many computers, typically in parallel, but not always Local Grid Computing
Ballou OpenCUDA+MPI
Applications of Supercomputing What can we do with Parallel and Distributed Computing?
Solving (Large) Linear Systems LINPACK Benchmarks
Fluid Dynamic Simulations N-Body Simulations Brute-Force Password/Hash Cracking Prime Number Searching Protein Folding Image Analysis / Manipulation ...
Ballou OpenCUDA+MPI
Who Uses Supercomputing?
ALL The People
Ballou OpenCUDA+MPI
Who Uses Supercomputing? No Really. . .
Google – Page Indexing Created Map-Reduce
Facebook – Data Mining Universities Many Others
Ballou OpenCUDA+MPI
1 GP-GPU Computing Introduction 2 Problems and Solutions
Problems with Current Solutions Solutions Plans and Goals 3 A Little About Methodology 4 What’s Been Done
Ballou OpenCUDA+MPI
Problems
”Distributed Programming” is expensive Specificity of Hardware Data Distribution Volume NFS
Fault Tolerance Optimizing Resources and Utilization
Ballou OpenCUDA+MPI
A Framework Solutions
Ease Programming Interface for Highly Parallel Distributed Computing Allow for Diversity in Computing Environment Bring together ideas from both types of distributed computing ”Jungle Computing”
Ballou OpenCUDA+MPI
Plan and Goals
Develop a framework for distributed computing over a heterogeneous cluster Develop several different solutions for vascular extraction from CT angiography scans Profile the different solutions Add Cluster/ Node Configuration and Scheduling Options Release as FOSS to the world
Ballou OpenCUDA+MPI
1 GP-GPU Computing Introduction 2 Problems and Solutions 3 A Little About Methodology
Implementation Details 4 What’s Been Done
Ballou OpenCUDA+MPI
Implementation Details
Arch Linux Salt Python CUDA (Open)MPI
Ballou OpenCUDA+MPI
Arch Linux
Core Tenet: Minimalism Small Lightweight Familiarity
Ballou OpenCUDA+MPI
Salt More than just NaCl
Provisioning tool for managing infrastructure Allows for ”Push” based state changes Remote Execution Simplicity Fast
Ballou OpenCUDA+MPI
Python
Development Speed: Expressive and Readable Fast Enough Written in C/C++
Great and Many Profiling Tools time.time() timeit cProfile ...
Where slow, allows use of C/C++ code
Ballou OpenCUDA+MPI
CUDA Compute Unified Device Architecture
Established interface with (nVidia) GPU’s pyCUDA Deferred CUDA kernel compilation
Familiarity
Ballou OpenCUDA+MPI
(Open)MPI
Established interface for inter-process communication mpi4py One of the most complete MPI implementations
Ballou OpenCUDA+MPI
1 GP-GPU Computing Introduction 2 Problems and Solutions 3 A Little About Methodology 4 What’s Been Done
What I have been doing Moving Forward Problems Encountered Potential Solutions
Ballou OpenCUDA+MPI
Tasks
Learning MPI, mpi4py, and pyCUDA Node/ Cluster Administration Node Build Scripts Salt Configuration Special Thanks to Danny
Lots of thinking
Ballou OpenCUDA+MPI
Moving Forward
Finish Creating Salt States and Configuration Continue Learning MPI and pyCUDA Develop Framework
Ballou OpenCUDA+MPI
Problems / Roadblocks
Power Requirements NFS Share /home performance Time?
Ballou OpenCUDA+MPI
Potential Solutions
Request(ing) more suitable and stable power Researching Distributed Filesystems / File storage
Ballou OpenCUDA+MPI
OpenCUDA+MPI A Framework for Heterogeneous GP-GPU Cluster Computing
The course covers data parallel execution models, memory ... PLEASE NOTE: THE ONLINE COURSERA OFFERING OF THIS CLASS DOES NOT ... DOES NOT CONFER AN ILLINOIS DEGREE; AND IT DOES NOT VERIFY THE IDENTITY OF ...
Feb 10, 2013 - This over develop- ment of wireless technologies using static resource alloca- .... 3. Energy-cognitive cycle. In this section we describe the structure of the proposed energy-cognitive ..... An alternative solu- tion can be used by ..
Flachaire E., Hollard, G. et Luchini S., Heterogeneous anchoring in dichotomous choice valuation framework,. Recherches ... the contingent valuation method in eliciting individual willingness to pay 1. In the dichotomous choice .... with a âmissing
which virtual âcreaturesâ compete for space and energy. We will ... the ability of evolution by natural selection to drive the increase in fitness of ..... of energies ϵ.
a type of application with multiple scales of communication. ... Chemistry and Chemical Biology, Rutgers University, Piscataway,. NJ 08854. â Electrical .... ity built on the BigJob/SAGA distributed computing envi- ronment ... Fortunately, great pro
Aug 10, 2017 - failure. In a consortium of banks, members could be large, global, systemically important financial institutions (GSIFIs). ... End users, such as a bank's customers, do not have an identity in the Coco network and cannot transact .....
Item 1 - 6 - Proc-2013-1-3-15.pdf (A visual DPF tool implemented by our students) ... Elements (in the signature) having the same name are considered ... {LOCAL PATH ON YOUR COMPUTER}/no.hib.dpf.text.updatesite/target/site. 8. Open the ...
Erlang'10, September 30, 2010, Baltimore, Maryland, USA. Copyright c 2010 ACM ...... rate on a budget dual-core laptop was 500 requests/s. Using parallel.
Prepared for the U.S. Department of Energy, Office of Electricity Delivery and Energy Reliability, under Contract ... (ORNL), and the National Renewable Energy.
Introduction to Framework One [email protected] ... Event Management, Logging, Caching, . ... Extend framework.cfc in your Application.cfc. 3. Done. (or in the ... All controllers are passed the argument rc containing the request.context, and all v
approach helps to predict QoS ranking of a set of cloud services. ...... Guarantee in Cloud Systemsâ International Journal of Grid and Distributed Computing Vol.3 ...
Sep 24, 2014 - potential network failures (N-1) ... in one of these roles in order to ... Users can't find out about data/services they don't have access for ...
Attribution 4.0 International License. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. Build. Assess. Iterate. Certification. Publish!
Keynesian multipliers should be big in liquidity trap. Crude Keynesianism: .... Calibrate income process to match macro and micro data. 2. Calibrate other ...
As will be seen in this section, images suffering from speckle noise should not be treated with the usual additive-noise derived tools (Wiener filter, for instance), since speckle corrupts the signal in a ... its estimation is discussed in [17] and [
Nov 2, 2012 - with higher prevalence and increases in medical care service prices being the key drivers of ... ket, which is an economically important segmento accounting for more enrollees than ..... that developed the grouper software.
needed to express one aspect of one per- cept or another. .... to layer 1. Drawing from de Lima, A.D., Voigt, ... permission of Wiley-Liss, Inc., a subsidiary of.