Re: [atlas-devel] 3.9.72

[Clint W. <wh...@cs...>]

>I'm getting back to testing ATLAS (and R using ATLAS) after a long
>hiatus. I just did a build on my laptop of 3.9.72 with full LAPACK and
>it looks like it's peaking at 17 GFLOPs (openSUSE 12.1, /proc/cpuinfo
>attached.) This is with the default flags - I don't see much point in
>trying to game it when I'm getting that much horsepower out of a $600
>machine and it only takes a few minutes to do the install. ;-)
>
>The bad news is that I want to run this in a virtual machine. I have
>kvm and xen on my Linux machines, and both VirtualBox and VMware
>Workstation 8 on Windows. I'm hoping eventually to extend this to an
>Amazon EC2 instance and an OpenStack compute node. What's a reasonable
>strategy for building images? I don't know what the low-end processors
>are in host servers these days. They pretty much have to be 64-bit
>just to get the Intel or AMD virtualization bits, but other than that,
>how bad can they be / how generic a build do I have to make?

My ESP is not bringing in the picture clearly.  Remember that we *do*
now have the ability to build generic defaults for pretty much all
hardware by manually setting the architecture to one of the new
generic (x8664x87,x8664SSE1,etc.), and manually setting the ISA
extension to use (SSE1,SSE2,SSE3).

If you want something that runs on every 32-bit machine ever produced,
you pick x8632x87; this uses x87 unit, and in the worst case (old intel
hardware) will reduce your peak by 1/4 for single and 1/2 for double.
I don't think any 64-bit machine ever lacked SSE1.

The way I guess I would approach it is to run the ATLAS configure process 
on as many instances as you can, and see what the least capable architecture
that is ever detected is.  Or perhaps you could ask someone what the
guaranteed ISA compatibility is?

?
Clint

**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************