A Cyberinfrastructure platform to meet the needs of data intensive radio astronomy on route to the SKA

SKA Calim 2011 - Day 2

Square Kilometre Calibration Meeting 2011 at the University of Manchester, UK.

Day 2

Next generation Calibration: John Q. Astronomer's Perspective (Ian Heywood) - Primary beam issues at low decs. Calibration in CASA. Intrinsic or apparent transients? Either way this can make a mess of your continuum map. Direction-dependent calibration... peeling. AIPS task PEELR... casa tas called peel but can't find any where apart from in the documentation. Followed Smirnov A&A, 572, 107, 2011. Didn't work, so used differential gains as a solution. (same paper). Solve for differential gains with MeqTrees, subtract model and image residuals. Used Offringa et al., MNRAS, 4 e05, 155, 2010 for the flagger (rficonsole). No phase switching as this was the lofar data - need to deal with them separately for each field - decided the cal sources were noise.

Fringe Fitting (Stephen Bourke) - Standard techniques; baseline based, FFT the visibility data - locate peak in delay, fringe rate space; Alef, Porcas method - baseline with closure constraints;  Schwab, Cotton Method - global fringe fitting - what happens in AIPS. CASA implementation - highly object orientation. Have implemented this inside of CASA. Hope for a BETA version in the next Casa release.

Solving for primary beams, pointing errors and the WSRT wobble (Oleg Smirnov) - cleaned up image to 1.6million:1 dynamic range - how to get this - apply differential gains. Requires heavy computing - the solutions scales as N**3. Main difference between this and peeling is this is simultaneous for all sources. Pointing slef cal doesn't seem to improve the maps so much - probably due to an inadequate primary beam model. QMC project - pick a field containing a cluster of reasonably bright off-axis sources and deliberately introduce pointing errors.  Actually found an error in the telescope pointing in the case that there shouldn't have been any offset. Independent solution on each band. Differential gains work better than the pointing solutions. The wobble is a 20 minute sine wave. No explanation yet.

Complex Factor Analysis (A Mouri Sardarabadi) - General likelihood ratio test can be used to see which model of the beam should be used. Approach is eigenvalue decomposition, so like principle componenet analysis.

LOFAR: Imaging Pipeline and MSSS (George Heald) - status of pipeline. First part does flagging and substraction of bright sources. Global sky model used to produce a local model used for calibration (need todo comissioning survey). Source extraction used to update global sky model. Level of rfi is low, some flase positives (in the LBA). HBA a bit worst. Demixing is used instead of direction-dependent gain solutions (van der Tol et al. (2007) IEEE TSP, 55, 4497). Good models of off axis sources are reuqired. Demostrated in python, but a little slow. With demixing + cleaning (without proper lofar beam models) the thermal noise is reached! Self calibration thus (no beam models) does not work well at the moment. Possible wrong dipole configurations on some antennas, i.e. they were putting in the wrong way around. Approx 30min todo a 6 hour dataset for the calibrations. As models get more complex this will take a longer, something like N**3 in the cases of the direction-dependent gain soltuions. Using new imager awimager - part of CASA as Sanjay mentioned.  Some other source finding routines.  For a 6 hour observation the demixing takes ~5hrs. LOFAR imaging cookbook - TAKE A LOOK.

LOFAR RM synthesis pipeline (Anna Scaife) - description of RM synthesis. In general case Faraday depth does not equaly the rotation measure. LOFAR low band can only observe small rotation measure values but get great resolution. e uses FFT based synthesis, from u,v,lambda cube to ra, dec,lamba cube.output for a pulsar gives RM of -61 when should be +61 suggestion is that LOFAR dipolse are the wrong way around. Possible wavelet based RM synthesis.  Compressed sensing can recover thin sources, Li et al. 2011. (Question: appears to have similar speed to GALFACTS).

Scaling for ASKAP and SKA1 (Tim Cornwell) - need 10,000 cores for ASKAP. Currently using 8800 cores on a machine called EPIC. Computing number of corse is about te scaling curve as targetted. SKA will require something like a billion cores. Build on top of casacore but don't use any of the casa imaging code. One MPI process per core, eventually will move to one MPI process per node - openMP for processing within a node then.See SKA memo 128. Latest machine uses the Intel Many Integrated Core (MIC) beta program - only 25 in the world at the moment. Very good results, can't show results. Discussion of snapsot imaging. Discussion of scaling of the imaging algorithms to cope with the scaling. SKA1 easy compared ASKAP. Data volume: spectral line ~ baseline; continuum ~ B**2. Computing in 2018 - the mean time to crash will be something like 1 day and it will be too expensive to write out the save state! Expect to move away from measurement sets.

Recent Widefield Wideband imaging with the EVLA (Urvashi Rao) - low signal-to-noise- discussion of accuracy of spectral-index vs frequency-range. Extended emission results: arxiv:1106.2796