Page dedicated to demonstration of HPGridding cases and profiling performance.
Tests performed by SCG are a demonstration of the effectiveness of the GPU gridder vs the standard CPU gridding code and should be viewed as a gate prior to verification and validation by CASA and stake holders. SCG tests will focus on questions of functionality (i.e. does it work at all, is the image scientifically plausible), performance impact of various test parameters (e.g. number of w-planes, image size, gpu model etc) and performance comparisons to standard CPU gridders.
Phase1 (VLASS-esque imaging)
Time Frame: T+5 weeks from receipt of code with w=32 support
Goals: Demonstrate the GPU code can be linked into full imaging (PSF, PB creation, gridding, deconvolution). First approximation of expected performance gains vs CPU based imaging
Caveats: Initially the GPU gridder will only support NTerms = 1, separation of PSF and PB creation is not possible with tclean so the htclean() distributed imaging task will be used.
Data sets: VLASS 2.1 11 square degrees from T10t34 CNSS tile
Process: Serial CPU+tclean vs 12 way parallelization of CPU+tclean() vs GPU+htclean() using awproject, nterms=1, initially with wplanes=1 and then with wplanes=32 when available, other parameters pulled from VLASS SE imaging workflow
Results: Compare images, characterize differences, compare runtimes characterizing serial CPU vs GPU performance and total system performance, extrapolate to expected scaling for full VLASS imaging with nterms=2 and full imaging workflow.
Phase 2 (VLASS imaging)
Time Frame: TBD June/July - August/September
Goals: Demonstrate CASA implementation (pre-release of 6.3) of full VLASS imaging, requires separation of per Taylor Term, PSF and PB imaging. Produced images could double as verification for SRDP. Validation would require a released pipeline which natively calls the GPU gridder.
Caveats: None
Data sets: Presumably reproduction of the 100sq dg imaging or some subset targeting high/low dec fields where AWProject is expected to be necessary.
Process: Full SE-Imaging workflow execution using GPU gridder. May begin as a hybrid tclean/htclean and morph to pure tclean as functionality becomes available.
Results: Work with SRDP for image comparison, characterize performance gains, generate full imaging estimate based on mix of CPU mosaic + GPU awproject imaging for all VLASS.
Phase 3 (other imaging)
Time Frame: TBD, can overlap Phase 1/2
Goal: Demonstrate non-VLASS imaging cases.
Caveats: If it's implemented we can test it
Data sets: TBD
Process: TBD
Results: TBD