[daip] aips on linux/amd64, multihost site
Eric Greisen
egreisen at nrao.edu
Wed Oct 31 10:20:35 EDT 2007
Vincent McIntyre writes:
> We have some sparc and some i[3456]86 machines, in separate sites
> (because of the endianness differences in the AIPS on-disk data formats).
> All share the same $AIPS_ROOT, otherwise. This works well.
>
> We are now considering putting in some amd64 machines.
> However the install.pl script appears not to distinguish i386 and amd64,
> in the sense that it sets $ARCH="LINUX" for both. So if I understand
> correctly, if one were to try to build AIPS from source on an amd64
> machine, for the same site as the i386 installation, this action would
> overwrite the existing binaries for i386.
>
> The questions:
>
> 1. Is it a supported configuration to run linux i386 binaries on amd64?
> I'm expecting "no" but thought I should ask to be sure.
Actually this works fine.
>
> 2. Assuming we set up amd64 binaries in a separate site, would it be
> possible to access AIPS data files from both i386 and amd64 hosts,
> without corrupting the data files? The data files would be accessed
> over NFS.
The formats are identical including byte order. But separate
sites means no normal access.
>
> 3. While I was looking into this I noticed there was no simple listing
> of "supported architectures" on the website[1] aside from [2] which
> doesn't mention amd64.
> Is it worth reworking [2] as a bullet-list (derived from install.pl)?
> It might be better to break that list out into a separate file (eg [3])
> and just linking to it from the current release info file (ie [2]).
>
I don't really view the AMD64 as separate. The Linux binaries that we
ship are linked/optimized for 3 architectures and run faster on
amd64's than the older GNU executables even when built for the 64-bit
architecture. The latest gfortran seems to have made up this 7%
difference, but treating AMD64 as really different does not seem
necessary. My recommendation - do not separate the amds to another
site and use our binary install. I suppose that curtails local AIPS
software development, but do you still do that? I know you once did,
but we have worked to reduce the need for that I thought to zero -
unless you have new/different tasks not ported to us.
Eric Greisen
More information about the Daip
mailing list