[Beowulf] motherboards for diskless nodes

[Jakob Oestergaard]

On Sat, Feb 26, 2005 at 10:59:57AM +0000, John Hearns wrote:
> On Fri, 2005-02-25 at 10:31 -0800, Greg Lindahl wrote:
> 
> > 
> > Doesn't make any sense; I have seen people describe such systems where
> > they download a disk image when a batch job wants a different software
> > load. It's certainly doable that way: it does have different tradeoffs
> > from the diskless case, but if it gives you a headache, it's probably
> 
> I've always dreamed of using User Mode Linux images for this.
> In a Grid-based world, prepare a UML instance which has all the
> libraries and runtime to run your code. Ship it across the grid with
> your executable. 
> The cluster at the receiving end can be running any distribution - it
> runs your UML in a sandbox.

Please see RFC 1925, corollary 6a:
 It is always possible to add another level of
 indirection.

Coming from truth 6:
 It is easier to move a problem around (for example, by moving
 the problem to a different part of the overall network
 architecture) than it is to solve it.

Your UML is, as its name implies, a user-space application, just like
the real application you were actually trying to run.  If your
application cannot be run on a given distro, I pretty much doubt your
UML (which is a very very complex user mode application) will run.

What you want is KISS: Keep It Simple (Stupid)

Don't link to a gazillion libraries if you don't have to. Link the
libraries statically when feasible (gives you a performance gain in many
cases anyway).

A statically linked application, or one with only glibc linked
dynamically, will run on very wide ranges of distributions.

Trust me on this; I make a living from selling an evil capitalistic
closed-source solution which needs to run on a very wide range of
distributions (and no, we do not link glibc statically because we're not
allowed to, but we keep our dependencies minimal and our binaries do run
on a very wide range of distributions).

> 
> And before anyone says it, yes performance would be a dog,
> and I don't see how UML could access all those nice Myrinet and
> Infiniband cards. SO I'm definitely blue-skying.

Again; adding layers of indirection is rarely a solution.

-- 

 / jakob