investigate performance issues

[VincentMcIntyre]

We may hit some usability issues with processing CABB data.

A simple calculation shows that reading a 5Gb file at typical disk access speeds (100Mbyte/s) takes around 45sec. That's for one scan, from end to end. Faster disks (eg SSD) will only give factors of 2 to 3 more read speed. Write speeds will be similar for HDD, slower for SSD. RAID disks accessed over Fiber Channel are limited to a line rate of 2Gbit/s, so you only get about a factor of 2 at best.

It seems worth while to build a version with -gprof and try to find the routines that take the most execution time. Perhaps they can be optimised? Or even "memoized"?

The other approach that could be explored is automating the splitting of work across CPUs, giving a set of channels to each CPU in the system. This is likely hard to do well, to do it at the FORTRAN level may require major changes that would be out of scope for us.

Emil's ATLAS uv data set is a useful size, about 6Gbyte.