Re: [SSI-users] QM400 and open-ssi

[John H. <jo...@Ca...>]

Vincent Diepeveen wrote:
> If i add a quadrics QM400 driver (patch to kernel),
> is it possible then somehow to get to work open-ssi 1.92 over the  
> qm400 cards?
>
> When needed i'm prepared to do effort there, if i knew what.
>   

You should have a look at Stan Smith's OpenSSI over infiniband stuff.

In outline the work to be done involves porting the internode 
communications code (ICS, part of CI) to your Quadrics stuff and making 
a initial ram disk that sets up the communications.

Maybe try contacting Stan directly?

> Even if it would require first boot over tcp/ip. I'm interesting in  
> getting it to work somehow.
>
> My chessprogram Diep uses shared memory. Ideally each process  
> allocates its own shared memory and attaches to each
> other process the shared memory. Each process is basically  
> calculating within its own chunk of shared memory,
> so that shared memory must get allocated at the node where the  
> process was started.
>   
Shared memory "just works" on OpenSSI.

When a page is mapped read-write by a node then is is marked "absent" on 
all the other nodes.

The r/w node just accesses it as memory.

If another node tries to read it then it gets marked read-only and 
copied over to the other node as a read-only page.

If another node tries to write it then it gets copied to the other node 
as a read/write page and marked absent on the other nodes.

So:

1 writer, no readers - no problem.

many readers, no writers - no problem.

1 writer many readers - some latency.  Has anyone measured this lately?

> Sometimes processes reference shared memory at remote nodes. The  
> important thing is that open-ssi gets that reference from
> the remote nodes over the QM400 network and not over the tcp/ip  
> network. Latency is important. Then i can get world champion
> computerchess next month in China with quadrics.
>   
Next month!
> I see by the way it has kernel 2.6.11. This might be a problem unless  
> when this is a vanilla kernel.
> I see quadrics has support for vanilla kernel 2.6 to integrate its  
> driver into the kernel (with a source code patch and
> building modules after that).
>   

Well, it's obviously not "vanilla", it has the CI and OpenSSI code!

To see if the patch applies suck it and see,
> Especially i see it support vanilla 2.6.18 kernel.
>   

No ETA on 2.6.18 at the moment.