Observations taken on long baseline optical interferometers are dramatically degraded by large (~100 wavelengths) phase offsets introduced by the atmosphere which fluctuate by wavelengths on millisecond time scales. The basic effect is to convert the improvement in target phase measurements from root-time to essentially no improvement at all, as the fringe phases average to zero.
A new technique just introduced (using the frequency dependence of the atmospheric index of refraction) allows the total phase offset to be partially removed over time intervals substantially longer than the usual atmospheric coherence times, dramatically improving the resulting S/N. The algorithm basically takes the time series (~few times 10^4 frames) breaks them up into chunks of 20, and processes the chunks identically. (I understand this is called an "embarrassingly parallel" problem). Development of the algorithm was done on an array at Los Alamos where, according to my colleague: "Processing one scan of 44000 frames in the fitting algorithm now takes about 4-5 minutes when I process 19 frames together with constant atmospheric phase and quadratic atmospheric delay over those 19 frames, and running in parallel on 200 processors".
I am currently applying for an NSF grant to support my observing effort at the Navy Prototype Optical Interferometer (NPOI) and would like to indicate that the campus could support the computing effort for our observations. The SeaWulf cluster seems well suited for this purpose.
Our needs will actually be modest. This instrument (augmented with these reduction techniques) are probing such absolutely untouched areas, that each object observed is essentially a paper - which takes a lot longer to write than the data reduction time. So our initial request would be for only 120 wall clock hours.
As indicated in the summary the algorithm involves estimating the total optical path difference (OPD) (fringe phase) introduced by the atmosphere to within a wavelentgh by looking at the differential refraction effects. The NPOI observes in up to 32 wavebands covering about an octave of frequency, 850 - 440nm, allowing this to be accomplished. Groups of 2 millisecond scans can then be combined, using a simple model (low order polynomial) of the evolution of the OPD during the time interval, allowing a corresponding ~5 fold improvement in the S/N. These chunks of visibility and phase data can then be combined over longer intervals from which we extract the amplitude and phase of the complex visibility. More detail on the problem and its solution can be found in a paper: Jorgenson, A., etal 2005, Proceedings of the SPIE Meeting on Interferometry in Optical Astronomy, Vol 6268, paper # 54, 2006