[rescue] SGI, Origin 200, and speed?

[Charles Shannon Hendrix]

Thu, 29 Apr 2004 @ 18:22 -0400, Dave McGuire said:

> >At NASA, Craylink was a bottleneck and the SGI rep said there would
> >be a replacement at some point.
> 
>   A bottleneck...?  As I understand it, it's a direct port off of one 
> of the crossbar switches.  It runs at memory speed.  How was it a 
> bottleneck?

It is always faster to access memory that is physically on the local
node.  Irix tries to make sure that threads are very close to their
memory addresses.  It isn't perfect.  It is also harder to make this
happen with a single-threaded application than a parallel application.
The performance papers from SGI say to keep memory access local whenever
possible and this is true of most any network memory interconnect
system.

The Origin will function as a big SMP box, but that isn't the best way
to use it.  The SGI code performance papers say to minimize Craylink
traffic and give guidelines on how to optimize.  You also have library
routines like dplace() which let's you bind memory accesses to the same
node the thread runs on, etc.

NASA insisted on code not being changed to make it optimized for SGI,
and SAIC claimed that the application could not be made parallel.  So
it ran as a big SMP application.  Fragmented memory accesses across the
cluster are much slower than keeping them local.

However, NASA felt that optimization would make it hard to move to
another system.  Therefore they used the Origin as a big SMP system.
They used Sun Grid Engine to keep all 32 nodes in the system busy,
without overrunning memory (36GB).  

A faster Craylink would have made their applications run faster.

Of course, so would writing parallel code, but that didn't happen.  I
think it was politics.

> >I figured if they did that, they might make it available for older
> >machines, and that you might know about it.
> 
>   I've not heard of any such thing, but then I'm not up to speed on 
> their absolute newest stuff.

The Origin 3000 website talks about Xtown2 ports as being bidirectional
1.6GB/sec interfaces, which is faster than Craylink.  It seems to be
used for I/O though... in fact I'm reading the site now and don't see
where it specifically says what is used for the interconnect fabric.

Unlike their O2000 pages which lists the memory bandwidth, the O3000
site only lists the aggregate bandwidth.

Need more info.

Man, I'd like to work on one of these.



-- 
shannon "AT" widomaker.com -- ["There is no such thing as security.  Life
is either bold adventure, or it is nothing -- Helen Keller"]