Re: [SSI-users] Can OpenSSI Be Used for General-Purpose Computing?

[Robert W. <no...@ro...>]

On Mon, Sep 01, 2008 at 12:17:37PM +0200, John Hughes wrote:
> Will Keaney wrote:
> > Hello.
> > I came across your project while doing some research on Linux 
> > clustering.  One question that I don't find clearly answered in the 
> > available documentation is, can unmodified applications be run on an 
> > OpenSSI cluster?  The other clustering solutions I have found all 
> > require specially modified code to use their libraries (Beowulf, PVM, 
> > et al), and I am really looking for something that is more versatile.
> An OpenSSI cluster tries to look as much like one big multi-processor 
> machine as possible.

As far as I know, with one big difference:

On a multiprocessor machine the OS' scheduler chooses automatically,
on which processor a process will run.
In OpenSSI the default is, that a process (with all of its threads)
runs on the node, where it is started, unlesse you say explicitely
otherwise (e.g. started in a "bash-ll", your application is made known
to be able to migrate or so ...).

Or is that not correct (any more)?


> The limitations are all related to the (much) higher latency/lower 
> throughput for inter-node memory access.
> 
> So:
> 
>    1. Cross-node shared memory is slower than normal shared memory.
>    2. Other forms of cross-node IPC are slower (Unix domain sockets,
>       SysV semaphores, SysV message queues).
>    3. Multi-threaded applications run with all threads on one node.
>    4. Accesses to filesystems mounted on other nodes are slower.
> 
> If your application can be written as many mostly independent processes 
> then OpenSSI should be very easy to use and give acceptable 
> performance.  (A good example: Linux kernel compiles with "-j" set to 
> number of nodes times number of cores).

Ack.


Regards, Robert