Research Plan (HHMI) - Summer 2004

Folding@Clusters

• Initial release with support for one MD package
  • Alternative to LAM-MPI

• Molecule testing and benchmarking
  • GROMACS
  • One other MD package

• Resource and software characterization
  • By molecule
  • By method
  • By platform

  • One specific item in this area is a comparison of system resources for a given molecule and method on PPC vs x86. On PPC there seems to be a lot more system overhead, could this be it?

    The interrupt load from a single GigE can be high enough to saturate the CPU, thus the introduction of NAPI in more recent Linux kernels, as well as Jumbo packets.

• Plan
  • Design, write, test, and release the code.

• Poster for submission to _____________?

Numerical Methods for Modern SIMD Architectures

• recip(x) and sqrt(x) are the principle operations of interest.

• AltiVec, SSE, possibly SPARC are the architectures to focus on.

• Questions
  • Which algorithms does GROMACS currently use for plain, AltiVec, and SSE?
  • On which platforms is it faster to do recip(x) in hardware rather than software? sqrt(x)?
  • Would a fused multiply and add instruction be useful in this context?

• Plan
  • NB: For now do all of the following on bazaar, not Cairo. John should use b16 and Dawit b17.
  • Get the tarball for GROMACS 3.2.1 and unpack it into a subdirectory of your home directory on the cluster file system.
  • Get the set of optimal configuration options for GROMACS and the information you need about FFTW from JoshH or JoshM.
  • Build GROMACS in a subdirectory of your home directory, do not do a "make install", just use the binaries that are built in your directory.
  • Run GROMACS with the villin molecule on a single processor. For most of the numerical methods work you won't ever run in parallel mode. Either JoshH or JoshM can point you to the molecule repository. Make a copy of the molecule source files to a subdirectory in your home tree before you start.
  • Look at the output of mdrun, compare the run-time data with (Dawit || John) and make sure you have similar results.
  • Repeat the above for two or three other molecules. JoshH can give you pointers to ones that will run in a reasonable (less than 1 day) length of time.
  • Put together a simple chart that shows molecules, wall time and CPU time for a single processor run, and for each molecule a list of the routines which taken together account for 80% or more of the total M-Flops consumed (with M-Flop and % data values). For instance if 3D-FFT takes 82% for villin than it would be the only one listed. If 3D-FFT takes 60% and the next smaller one takes 22% they would both be listed, etc.

  • Build a benchmarking kernel based on GROMACS for these operations.

• Paper for submission to MCURCSM, _____________?

Benchmarking and Tuning GROMACS Paper

• Choose an outlet

• Write the paper

• Test GROMACS with PVM and other MPI implementations and report on the results.

• Consider adding a chapter on tips and failure modes for running in parallel mode.

Plumbing - see http://cluster.earlham.edu/project/sna/plumbing-summer-2004.html That document is a candidate for a WiKi.

People and Project Mapping

• Charlie - grants, plumbing, F@C, b-and-t gromacs paper,
• JoshH - F@C, b-and-t gromacs paper,
• JoshM - grants, plumbing, F@H, b-and-t gromacs paper,
• Dawit - grants, plumbing, numerical methods,
• John - grants, numerical methods,

todo

• consider wiki for items where it's appropriate