Re: [Linux-aus] Grant request

[Davyd Madeley <davyd@madeley.id.au>]

On Sat, 2005-08-13 at 17:24 +1000, Michael Still wrote:
> Davyd Madeley wrote:
> 
> > The machine manbo has been upgraded with fast disks and lots of RAM
> > (around 11G), what we want to do now is upgrade the CPUs. For AUD$1300,
> > plus shipping, we can order 20 new CPUs for this machine, increasing its
> > processing power by a factor of around 4x.
> 
> I looked at the UCC web page for the machine, and it's currently chock
> full of slow CPU goodness. Could you provide information about:
> 
>  - the speed of the new CPUs

400MHz Sparc64. We plan to order anywhere up to 20 of them. The machine
will take 16 IO boards (each will handle 2 CPUs). The only caveats on
this are the CPUs must be matched (although they can have different
amounts of cache) and that some of the IO boards are required to do
things like talk to the disk array.

>  - are they new? Warranty?

Unlikely. Warranty may be negotiable. The dealer selling the parts is an
American company we have dealt with before to get parts (power modules
and RAM ingots) for this machine and another like it (but smaller)
previously.

>  - how much time will it take to perform an average regression run [1]?

Empirical testing indicates that it will not be waiting for work, the
tree is simply too large. If it is found to be building too often to be
useful (that would have to be a complete build every hour) we'll simply
load it with more testing tasks to complete. Some goals we have been
discussing include upgrading to Solaris 10 to use the new dtrace tool
and to allow dtracing of the desktop components. At the moment the build
is taking too long in my opinion. We should be able to get it down to a
couple of hours, my goal is to do it in under an hour.

>  - would this sort of project benefit more from donated CPU time on
> something[2] like the spanking IBM hardware that just got loaned to ANU?

There are four main points to running these builds on the sparc:
 - to quickly find out build regressions;
 - to identify build issues in building with high levels of parallelism;
 - to identify build regressions on Solaris/Sparc, which won't always be
obvious to the x86 orientated community; and
 - to utiltise new Solaris analysis tools such as dtrace

> 1: I know nothing about tinderbox, but I want to know if your machine
> would always be behind reality, or always waiting for now work.

To build the entirety of GNOME on my old laptop (an old pentium 4) took
13 hours. To build it on my old desktop (a newish pentium 4) took 6
hours. I haven't tried it building the stack top to bottom on my new
laptop or desktop (no time). Manbo is currently taking around 11 hours,
it has fast disks (Seagate Cheetahs, I think) and lots of RAM, so what
we need is a CPU speed boost.

> 2: That is, I'm raising the concept, not commiting the ANU machines, as
> that is outside my power.

Understood. I am also developing some software to parallelise builds on
an Alpha/Linux cluster. This is however going to be a much more
challenging project. It will also not give us the ability to utilise the
Solaris features I've mentioned above.

Kind Regards,
--davyd madeley
UCC Secretary, 2005

-- 
Davyd Madeley

http://www.davyd.id.au/
08B0 341A 0B9B 08BB 2118  C060 2EDD BB4F 5191 6CDA