Parallel Network-RAM

Large scientific parallel applications demand large amounts of memory space. Current parallel computing platforms cannot handle the memory load of large scientific simulations without degrading performance. To solve this problem, we propose a new solution called Parallel Network RAM. This approach will allow larger problems to be solved while minimizing the computational, communication and synchronization overhead typically involved in parallel applications.

People

Advisor ~ Li Xiao - lxiao@cse.msu.edu - home page
Graduate Student ~ John Oleszkiewicz - oleszkie@msu.edu - home page

Project Documents

Summaries
  • Fall 2003 Final Report - pdf
  • Fall 2003 Midterm Report - pdf
  • Summer 03 work summary - doc
Presentations
  • Poster - (5/10/04) - ppt
  • Master's Thesis Presentation - (4/26/04) - ppt
  • CSE 891 Project Presentation - (4/21/04) - ppt
  • MSGC presentation - (10/11/03) - ppt
Progress Reports
  • Progress Report I - doc
  • Progress Report II - doc
  • Progress Report III - doc
  • Progress Report IV - doc
  • Progress Report V - doc
  • Progress Report VI - doc
  • Progress Report VII - doc
  • Progress Report 3-22-04 - pdf

Documentation

  • PNRSim Technical Report - Documentation on the usage of PNRSim and the algorithms and models it implements. html ~ pdf ~ ps ~ bibtex entry
  • Code Documentation - Documentation of simulator code (for programmers) ~ html
  • Hardware Times - a collection of realistic hardware performance times (very sparse at the moment) ~ html

Experiments

The following tar files contain configuration files and workload files that define a series of experiments. They also contain a sample script that will run the experiments (with tweaking of pathnames). These files are provided so that work performed for published papers may be duplicated by others.
  • Expr0 - Experiment set 0. This is a preliminary set of experiments done on a lighter memory access and synchronization load. The results of these experiments will guide what experiments will be done beyond the base set on real memory and synchronization loads. Note that expected results are provided in two text files. Run with input5000.
  • 64 PE Experiments - This set of experiments was used in the master's thesis research. It test Parallel Network RAM on a variety of cluster setups - although each cluster has 64 PEs (nodes). It uses the latest configuration file parameters and tests six different variations of Parallel Network RAM.
    Note that the "backbone" configuration files are misnamed. Backbone8 should be Backbone4, Backbone16 should be Backbone8, etc.
    Use any of this set of inputs.
  • 128 PE Experiments - These experiments were used in the master's thesis work also and are identical to the 64 PE experiments except that now the cluster system has 128 PE's (nodes) instead of 64.
    The "backbone" experiments are correctly named here.
    Use any of this set of inputs.

Code

The following is development code. There is no guarantee that this code is stable and will not damage your computer in some way. Download and use at your own risk.

Data

Technologies

The PNRsim project uses the following technologies
  • Simpack - a free (GPL) simulation building toolkit.
    NOTE: Major changes were made to the Simpack code. These changes allowed easier compilation on Linux and better integration to the PNRSim project.
  • DOC++ - an automatic documentation generator for C, C++, Java, PHP, and others.

Tools

  • parse.pl - .gz - This script parses the output of PNRSim and puts it into a nice tab-delimited line suitable for insertion into a spreadsheet.
  • scale.pl - .gz - This script takes a .swf workload file and scales processor requests down by the factor supplied. So, for instance, if a job requests 512 processors, and we scale that down by a factor of 32, then the process will request 16 processors.
  • backup.pl - .gz - This is a simple script I wrote that automatically bundles the PNRsim code and uploads it to this webpage. Its very specific to my setup, but if it gives you ideas, great.

Acknowledgements

This research was made possible by a grant from the Michigan Space Grant Consortium
Up Home