Travis Metcalfe (NCAR)
description
Transcript of Travis Metcalfe (NCAR)
Travis Metcalfe (NCAR)
A TeraGrid Science Gateway for Kepler Asteroseismic Data
Collaborators: Joergen Christensen-Dalsgaard,Orlagh Creevey, Matthew Woitaszek
Wampler’s screwdriver
Schematic pipeline
input global search local analysis output
nl, Teff, L, R
parallelGA
SVDLM
M, Z, Y,
pulsation frequenciesand other constraints
optimal parameters, other model output
teragrid bluegene desktop workstation
ASTEC + ADIPLS
GA convergence
• Initial convergence driven by the crossover operator (first 20-30 generations)
• Subsequent improvement due to random favorable mutation operations
• Three out of four runs find the global solution within 200 generations (ran 300)
0.75 < Mstar < 1.75
0.002 < Zinit < 0.05
0.22 < Yinit < 0.32
1.0 < mlt < 3.0
50,000 CPU-hours/fit
Surface effects
• Incomplete modeling of surface convection zone leads to systematic errors
• Calibrated with difference between fit to “Model S” and BiSON radial modes
• Global fit to BiSON data, including scaled empirical correction for all modes
BiSON results
• Fit to 36 frequencies with l = 0-2 and constraints on temperature, luminosity
• Matches frequencies with scaled surface correction better than 0.6 Hz r.m.s.
• Temperature and age within +0.1%, luminosity and radius within +0.4%
Science Gateway
• Web interface to specify observations with errors, or upload as a text file
• Specify parameter values to run one instance of the model, results archived
• Source code available for those with access to large cluster or supercomputer
TeraGrid workflow
128-way parallel
100-300 gen/week
Community management
• Public users: anonymous visitors will see the status of current model-fitting runs and will be allowed to browse/download archived runs
• Registered users: KASC members who provide name, institution, and email can download code, submit single-model and model-fitting requests
• Gateway managers: approve new registrations, modify job and queue policy, and manage public archiving of the model-fitting results
Data release
• Model-fitting in progress: will be shown with KIC number, J2000 coordinates, requesting scientist, institution, resources allocated and job progress
• Completed model-fitting: available to registered users during the proprietary period of the data, then archived on the website for all users
• Single-model requests: archived immediately on the website for all users (to avoid duplicate runs), with choice of binary files in big/little ENDIAN
Open questions
• Should all of the model-fitting results be visible to all registered KASC members, or do we need to define subgroups (e.g. for KASOC)?
• When should the model-fitting results be visible to non-registered users? How should the timing of the public release be determined?
• Should there be limits on the number of jobs submitted to the queue by each user? Limits on the number of jobs running? Job priorities?