On Tue, 11 Nov 2014, Tim Jenness wrote:
> On Tue, Nov 11, 2014 at 8:39 PM, Tim Jenness <[log in to unmask]> wrote:
>
> >
> > In other news ccdexercise still fails when memory mapping is enabled. It's
> > going wrong in makemos where something is messing with the data arrays
> > (even if noscale and nozero are used). The input files to makemos are the
> > same so ccdexercise is getting that far without having any problems.
> >
> >
> And now I know why. memory mapping a sliced locator is obviously a bad idea
> in many cases. Disabling all sliced locators (rather than attempting to do
> a check for contiguousness in cases like vectorized locators [which would
> also require me to calculate the offset into the data array]) allows
> ccdexercise to run. I was getting worried a bit there.
I don't understand the details here (haven't been following very closely)
but if ccdexercise/makemos is doing something inadvisable it's very
likely my fault since I had my grubby hands on CCDPACK early on in
my starlink career, i.e. before I had a good idea what I was doing -
if so apologies.
> Preliminary timing test:
>
> ccdexercise (real / user / sys )
Also from the mists of time: the very first thing that Rodney put me
to work on was assessing use of various starlink packages for "large"
images. One output of that work was a (sort of) benchmark script based
on ccdexercise for large images. This was called CCDBIG and documented
in SSN/69. That never made it into the USSC for reasons I can't
remember (maybe it wasn't very good?) but there's a copy here:
http://www.star.bristol.ac.uk/~mbt/docs/ssn69.htx/ssn69.html
I have RCS files for the document and the software, but I don't
know how much effort would be required to get them built.
If you think this would be useful for benchmarking the HDS5 stuff,
I could try to dig it out. However, given the comparison between
how much you know about starlink/hds now and how much I knew then,
it's quite likely that you'd be better off writing anything along
those lines that you need from scratch.
Mark
--
Mark Taylor Astronomical Programmer Physics, Bristol University, UK
[log in to unmask] +44-117-9288776 http://www.star.bris.ac.uk/~mbt/
|