Some relevant info...
Many scientific apps do not scale well beyond 32 CPUs/cores, or so
I was told during my training course for Cray systems. It depends
on how well a task can be parallelised. Sometimes a single shared
memory system like Altix is far more effective when data access
patterns are very mixed & complicated between nodes and RAM,
other times a cluster makes more sense when a task can be easily
split into smaller units. Lots of other issues aswell - cache/RAM
behaviour, data I/O requirements, etc. (eg. ANSYS runs often
need massive amounts of RAM but not many CPUs).
Some tasks marry better onto different architectures such as
Vector machines. It varies.
It's much more likely that the Jaguar system is used to run
multiple different jobs at the same time, partitioned into
numerous separate OS instances (some variant of Linux I expect).
Typical applications include cosmology, weather modeling, QCD,
CFD, GIS, and yes modelling nuke explosions, though actually
other common tasks for that programme include things such as
modelling corrosion, wear & tear, etc. - what happens to the
casing after 30 years? Is it still reliable? Problems such as
how to design the electrical systems so they can withstand the
radiation, circuit redundancy, simulation, that sort of thing.
Modeling a bang is a small part of the whole approach.
In reality, 80% efficiency on systems such as this is a solid
achievement, if possible. Usually though the logical approach
is not to even try and run codes which would be better suited
to a different architecture, or just use it to run multiple
different tasks.
I've always had a soft spot for shared memory systems (because
they can handle just about anything) but it's hard to scale
them beyond a few thousand CPUs and cost-wise they're not the
best choice for tasks which can be easily split up such as
rendering, unless custom sw is written to exploit the
architecture properly (which is what ILM used to do with its
32-CPU Origin2000 systems). Still, it's no less fun watching a
32-CPU system doing a Blender render.
I doubt Jaguar's top spot will last long though.
Also, lookout for SGI's new UV system being announced soon (AFAIK,
it's an i7 XEON NUMA shared-memory design - I hope so anyway),
and meanwhile we should start seeing large Nehalem clusters
appearing which ought to boost performance by a huge margin
given the big improvements available over the older XEON series.
Ian.