cannot download atlas3.9.67.tar.bz2 from any mirror

[johnnyangle]

Hi. Trying to use nix-1.3 to install octave-3.6.2 from source to local user account on an 64-bit Red Hat workstation. Dies with failed attempt to build atlas-3.9.67. I'll paste in the whole log file if requested, but this is the gist of it:

trying http://prdownloads.sourceforge.net/math-atlas/atlas3.9.67.tar.bz2
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
^M 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0^M 0 0 0 0 0 0 0 0 --:--:-- --:--:-- --:--:-- 0
curl: (22) The requested URL returned error: 404 Not Found

...tries multiple other mirrors...

error: cannot download atlas3.9.67.tar.bz2 from any mirror

[vcunat]

You should get the file from hydra automatically.
downloading http://nixos.org/binary-cache/nar/18vmn38y7laadjyiddp1dp5vgi6zqixyn6lmk7ky68ishlxd78xg.nar.bz2' to/nix/store/j9x7fhk66ks3rrvldic43awjgigivg5l-atlas-3.9.67'

Adding the channel should fix it (and likely save much other building).
$ nix-channel --add http://nixos.org/channels/nixpkgs-unstable
$ nix-channel --update

[vcunat]

But the real problem is deeper. From what I just found out, atlas wants to do machine-specific optimization in build time, which doesn't go very well with our build system -- purity, optimization for the builder and not the runner, etc. Now the upstream even forces the builder to switch CPU governor. They took out the option and, sincerely, it seems that using atlas in nixpkgs isn't a good thing (at least in hydra-built default versions of packages).

I see no quick solution to this, just switching to other implementations doesn't help because of some hard-coding in the expressions, perhaps someone will fix it in future...

[viric]

On Mon, Feb 04, 2013 at 05:32:57PM -0800, Vladimír Čunát wrote:

But the real problem is deeper. From what I just found out, atlas wants to do machine-specific optimization in build time, which doesn't go very well with our build system -- purity, optimization for the builder and not the runner, etc. Now the upstream even forces the builder to switch CPU governor. They took out the option and, sincerely, it seems that using atlas in nixpkgs isn't a good thing.

I see no quick solution to this, just switching to other implementations doesn't help because of some hard-coding in the expressions, perhaps someone will fix it in future...

For me this atlas issue is nothing new. But I'd wait until someone has a problem with it. :)