Cluster Computing Group

Brief Meeting Minutes - November 30, 2004

JoshM and Charlie met briefly, the current list looks like:

JoshM

• Install and try-out NAMD on bazaar and cairo.
  
• Build a new bazaar image (with Dawit) using SuSe 9.1
  
• Build a new cairo image using Yellow Dog 4.0 (including ccache and distcc).
  
• Continue working on F@C code items.
  
• Work with John on testing and the MidWife.
  

John

• Complete testing matrix using a1 for all molecules (GMXBench plus others), bazaar and cairo, 2-12 processors.
  
• The MidWife.
  

Charlie

• Update prose for two SIAM CSE poster submissions. (done)
  
• SIAM CSE conference registration for JM, JS, CP; JH are you going?
  
• Hotel and air reservations for SIAM CSE

• Organize F@C a2 release.
  

Work Log

• The MPI_Recv to MPI_Irecv problem highlighted int the mother is a non-problem. This is expected behavior between the mother and child. This can be taken off the ToDo list.
• I added the ability to correctly detect the number of CPUs on a Linux machine in COSM. i tested on Bazaar and Cairo.
• I cleaned up the "step" messages generated by mdrun. These should be collected by the nanny and transmitted to the mother for a status display.
  Finished Step 10 of 1000, remaining runtime: 90 s

• Created a new feel for the website. Let me know what you think, and if I should push it.
  See the link in the email, let me know if you need it again [I don't publish that private work space to keep it from being sniffed by spiders...]

Meeting Notes

• Short term code fixes (joshm)
  • Divorcing child and mdrun.
  • diff the configure.ac, makefile.am, and (?) to see changes and to preserve our FATC flag.
  • See if altivec issue is patched in the latest version of gromacs. If not, fix it.
  • gromacs version - is the highest production 3.2.1?
  • Don't do another code root; just roll back changes in CVS.
  • Changes to make in gromacs: remove new main in mdrun, linking instructions, base level Makefile.am will have to change (has dependencies on the child binary).
  • We need new build instructions.
  • Tabs and 80 cols.
• Run muliti-node tests on all platforms. (john)
• Update node usage doc with by giving all non-NFS nodes to John.

• Mount on board like the physics posters?
• Print poster content at appropriate size.
• Option: Laminate the poster?
• Two new generic framworks need to be created: FATC, Numerical Methods for SIAM
• MERCK - Tuesday the 16th at 6-9p(?)
• Dr. Dobbs nov 2005 issue - write article about FATC. Due in mid July. We need to start writing very soon.

• Test a1 binary on c12-c15. Run as a test user.
• Results table.

• Making the cluster space usable:
  • Ask John Howell about shelves.
  • Mount screen and projector.
  • Vacuum
• Data Visualization Console: temp file flooding and errors when running.
• Midwife write-up (Charlie)

• HHMI annual accounting and report (Charlie)
• Talk to Mary Lou about travel logistics for SIAM. Start a calendar, get details down, talk to Joshh.

Code Review Notes

• separate mdrun from child (like grompp)
  • Start with latest version of GROMACS.
  • keep define f-at-c (for stderr->stdout, verbosity of prints...)
  • Patch for altivec
  • Change build script
  • Take out include of gromacs header for f-at-c
• Remove HTTPD from mother
  • Turn this into MPI async communication.
  • Turn off Threads in COSM
  • No longer need to send the mother hostname and port to the nanny
• CHILD_MULT should be dynamic (load balancing)
• File System Recourses needs to be altered for dynamic files per pande stuff
• Documentation:
  • Header file has complete documentation for functions, and globals
  • Use special tags for start and stop of comments.
• Print function: Reconcile N-different.
• Remove debug counter [duplicate: debugCounter, debugTimes??]
• Functional abstraction in main() - mother/nanny/child
• MPI Send and Recv file function()
• Remove OpenLog()
  • Now we can use Print everywhere and should!
• Stdout redirection toggle [log | display]
• Route progress of child
  • caught by nanny
  • sent to the mother async. [disjoint from checkpointing mech.]
• Clear documentation of *all* variables.
• Trap, check, & react - return codes from all functions when appropriate.
• Use PrintHeader(x,x) in [-v] version return.
• Write a PraseSpec function to get applicable file names from assignment server.
• Pre-while loop set child_process
• Where is nanny_host_conf_file set in mother.c
• convert int -> %u to something reasonable (signed)
• If conf file is not set properly, loose default values pre spawn.
• When waiting for the child to return a result, should use Irecv instead of Recv
• check_code() function may be redundant.
• "Recover #" print messages, should be more meaningful.
• "hold" variable may be lost since it doesn't *normally* do anything meaningful.
• "flag" should be equal to some defines in the header. clarity issue
• Testing Rubic Item: Test stopping mother with a signal

Meeting Minutes - November 5, 2004

Folding@Clusters
Design handout for SC2004

Discuss legends for clusters on the diagram

Numerical Methods

Plumbing

Papers and Presentations

Items from JoshM's email of Wed