Re: mosix and debian amd64

[Paul Brook <paul@codesourcery.com>]

> >I've never understood the appeal of Mosix. No real HPC facility uses it.
> >
> >Proper clusters tend to use an MPI of some kind (usually some variant of
> >MPICH). Workstation clusters run Condor or similar. Neither require
> >special kernels.
>
> Hmm... I'm new to cluster calculations, so I don't know for example
> what does HPC stands for.
>
> As for MPI, afaik it is some library that has to be linked with a
> program and the program has to use some functions of the library.
> That means making the code dependant on the library.
>
> When using mosix the code doesn't have to be modiffied - it can be
> run either on a single machine (with no cluster specific software
> installed) or on a cluster.

I'd expect a decent batch submission system to do better than mosix if you're 
just running on a single node.

Compared to SMP systems, clusters have high latency and low interconnect 
bandwith. That same is true, to a lesser extent, of large NUMA systems. 
Unless you modify and/or recompile your application to be aware of this it's 
not going to scale past a single node. 

There are languages/frameworks (such as HPF or VISPL++) that make programming 
clustered systems easier, but you've still got to rebuild/rewrite your 
application to use these.

Paul