MultiGrid Interoperation TDDFT/Ninf-G Experiment

• TDDFT/Ninf-G Overview

• TDDFT Requirements

• MultiGrid Resources

• Account application
  1. The user had have a PRAGMA account before the Multi-Grid experiment.
  2. The user requested a TeraGrid account on the Web.
  3. The account was creaated on only UC/ANL site for the demo without the formal process.

• Deployment
  1. The user wrote a system requirement for the application.
  2. The GOC sent the requirement to the PRAGMA sites and the UC/ANL.
  3. After notification, the user confirmed the requirement is met on the site.
  4. The user upload an input data file (212 MB) to each site.
  5. The user upload an application program written in Fortran 90.
  6. The user compiled the program on the site.

• Test on each site
  1. Application level test
     • The user checked if the program works normally on a signle node.
  2. Local scheduler level test
     • The user tested a simple job submission to a local scheduler.
  3. Globus level test
     • The user tested a simple GRAM job submission from the client node to the cluster.
  4. Ninf-G level test
     • The user tested the TDDFT/Ninf-G/Globus application using the client node on AIST and a single node on the site.

• Experiments
  • The client always runs on ume.hpcc.jp (AIST).
  • Target molecules of TDDFT: Ligand-protected Au_13
    • 122 tasks in one loop
    • Input data per task: 4.87 MB
    • Ouput data per task: 3.25 MB
    
    
  • 2/10 10:34 JST - 10:56 (22 minutes)
    • 6 servers: 4 SDSC, 2 UC/ANL, 2 AIST
    • 1220 successful RPCs and 2 errors
    • The server programs on UC/ANL went down after 15 minutes because max_wall_time was not set by the user.
    
    
  • 2/11 00:06 JST - 2/16 01:13 (5 days and 1 hour and 7 minutes)
    • 14 servers: 6 SDSC, 6 UC/ANL, 2 AIST
    • 464162 successful RPCs and 6 errors
    • UC/ANL servers was queued for 3 and half hours in the beginning.
      • job_startTimout of Ninf-G was disabled.
    • The server programs on UC/ANL went down after 50-hours-running because of max_wall_time.
    
    
  • 8/16 17:44 JST - 8/17 02:11 (8 hours and 27 minutes)
    • 4 servers: 4 OSG
    • 1220 successful RPCs and no error
    • This was a test run on the OSG cluster and the result was successful.