UMICH-2015: Simulations & Data Analysis Break-Out Session 1

From CMB-S4 wiki
Revision as of 11:41, 22 September 2015 by Jdborrill (talk | contribs)
Jump to navigationJump to search

What needs to be in the S&DA section of the CMB-S4 Science Book?

For each mission phase (design, construction, deployment, observation):

  • What are the S&DA goals?
  • What (human, compute) resources are required? available?

Given this:

  • Where does the balance lie between veracity and tractability?
  • How do we ensure that our progress through phases is evolutionary not revolutionary?

Please add to the italic-headed lists of resources and their points of contact below



A hierarchy of methods of decreasing veracity/increasing tractability.

Simulations used for:

  • Forecasting/Design
    • sample mission model space for given sky model(s)
    • requires explicit metric(s), eg. Science/$
  • Validation & verification
    • representative & self-consistent mission realization(s)
    • bootstrapping V&V of both data analysis and simulation codes
  • Debiasing & uncertainty quantification
    • accuracy & statistics
    • dominant computational cost

For each simulation element (sky simulation, data simulation):

  • What is the current status?
  • What is required in each mission phase?
  • What are the challenges to meeting these requirements?

Sky Simulation Tools/POC

  • PSM (Delabrouille): IDL code generating 10-component foreground model
  • PPSM working group further develop the post Planck sky model - port into python?
  • Hydrodynamical sims for extragalactic secondary anisotropies - lensing, tSZ, kSZ, CIB, cross-correlations, point sources, etc. for forecasts

Data Simulation Tools/POC

  • PSM (Delabrouille): IDL code applying detector band-passes to sky model; pyPSM python re-implementation in progress as precursor to web interface
  • TOAST (Kisner): C++ MPI/OpenMP tools for massively parallel TOD manipulation, including OTFS of colored correlated noise timelines; pyTOAST python re-implementation in progress including OTFS of full-beam convolved sky timelines & diskless interface with pre-processing
  • CONVIQT/libCONVIQT (Prezeau/Keskitalo): full-beam convolution in the time domain
  • FEBeCoP (Rocha): effective-beam convolution in the pixel domain
  • Fisher matrix methods

Data Analysis


How do we combine data from multiple telescopes of multiple classes at multiple sites?

  • Working group needed here

What do we do about data covariance (functional forms, approximate matrices, Monte Carlos, other)?

For each data analysis element (mission characterization, pre-processing, map-making, component separation, power spectrum estimation, parameter estimation, ... ):

  • What is the current status?
  • What is required in each mission phase?
  • What are the challenges to meeting these requirements?

Map Making Tools/POC

  • MADAM/TOAST (Keskitalo): Fortran MPI/OpenMP destriper with TOAST OTFS & Monte Carlo capabilities.
  • Springtide (Ashdown): Fortran MPI/OpenMP destriper with TOAST OTFS & Monte Carlo capabilities.

Component Separation Tools/POC

Power Spectrum Estimators/POC

Parameter Estimators/POC

Data Challenge & Computational Resources

Cmb hpc scaling.jpg

Suborbital data growth exactly tracks Moore's Law; satellite data grow at half the rate.

Beware of committing to compression (lossy, outdated)

Computational Resources (Including Scale & Accessibility)

  • Currently available
    • NERSC: DOE general purpose HPC center with new top-10 system every 3 years; 1% of cycles annually for 20 years, accounts for anyone, public data repository,
  • Future potential
    • ALCF: DOE leadership HPC center with a new top-5 system every 3 years; limited number of users & 'heroic' computations, next-generation architecture in common with NERSC
    • NSC: Chinese HPC center with world #1 system; limited number of Chinese-led users & 'heroic' computations
    • SciNet: Canadian HPC center


  • Dedicated sky modeling activity!
    • Clean
    • Self-consistent
    • Usable
  • Pipelines & interfaces
    • Tightly-coupled pipelining where I/O is prohibitive (time-domain) - interface in memory.


    • Loosely-coupled pipelining where I/O is reasonable (pixel-, multipole-, parameter-domains) - interface on disk.


    • Support for both rapid prototyping and efficient production
  • Standard data objects & formats - combine with data working group
    • Define file and memory formats for interfacing, informed by computational efficiency.
    • Generalized mission model
  • Data distribution
    • Take-out vs Eat-in
  • Synergies with S3, LiteBIRD, COrE+, etc
    • Common sky models
    • Parallel pipelines (V&V, prototyping, general vs specific)

Fast mocks

Wiki navigation

Return to main workshop page

Return to Simulations & Data Analysis page