Re: [SSI-users] ssi on podcast

[Vincent D. <di...@xs...>]

>
> We did talk to the creator of ScaleMP, not open but this is what  
> our listeners asked to compare OpenSSi to: http://www.rce-cast.com/ 
> Podcast/rce-65-vsmp-scalemp.html
>

If i google for 'latency' ScaleMP, i basically find 0 links.

Having 0 public latency benchmarks whereas the most important thing  
of a SSI is the latency to remote nodes,
not how many 'tfop' it is, as tflops are not useful definition for  
CPU software that's latency sensitive.

The Tflop software that's doing the crunching in floating point  
world, that's all GPU's anyway that dominate there,
and they have a very bad latency to and from the RAM.

Huge bandwidth sure, but there is a huge difference between latency  
and bandwidth.

On the homepage of Scale MP there is nothing.

With spoken word you cannot easily confront someone with hard  
statements on what actually gets delivered.

So this is not an interesting hardware vendor as there is no simple  
to find information on the homepage showing
you the latencies at a given configuration and latency is everything.

It's not even clear if you read their definition whether they just  
deliver you a bunch of machines with software,
or just the software.

So i don't consider this serious.

Yet i'm sure some ex-SGI customers will love them.

If i may remind some people here how expensive all this was. Around  
year 2k you could get a 16 processor Sun or Dec Alpha
shared memory machine at a cost of $10 million.

For the same $10 mln you could buy also a huge supercomputer using  
NUMA, for example SGI.

The real clustering of common PC's became more popular when they won  
it from the specialized HPC cpu's. That's something
that happened past 10 years. Right now there seems to be a split  
between software that works well on CPU's and everything else
that can work on a GPU using gpgpu.

GPU's have won bigtime the HPC crown, so that leaves for the cpu's  
everything that's not easy to reprogram to GPU's and/or not
sellable as a gpgpu program.

It's not clear to me what scalemp delivers there from reading its  
homepage. Not quoting price is 1 thing, not quoting what you exactly
deliver is weird.