Top 500 trends
rbw at ahpcrc.org
Wed Nov 27 13:28:09 PST 2002
On 27 Nov 2002, Robert Brown wrote:
>On still other sides, "supercomputers are clusters".
>We should probably differentiate between vector supercomputers (Greg's
>point, which is well taken) and e.g. SP supercomputers, which are
>basically clusters. Really expensive clusters, but clusters.
This IS an important point ... in that Power 4 clusters even with
one core turned off do not compete with vector machines or even the
Alpha or the best P4 nodes on a stream bandwidth basis. So as "really
expensive clusters" as Robert puts it, what is the extra money going
for ... I guess excellent in-cache peformance, 64-address space,
some operating system niceties perhaps, and IBM's support/service
to a captured/installed market. There may be other items that I am
missing here, but most of these features can be had elsewhere for
a substantial discount ... they are just not Big or Blue.
On the other hand, the X1 and SX5/6 machines offer the still-unique
(but shrinking) per CPU bandwidth capability of the vector architecture
(now combined in systems that scale to 1000s of processors). The price
is high, but such a 1000+ processor, vector system (especially with
its BMM hardware) is the right choice for customers with ready supplies
of your tax dollars ( ;-) ).
They can even trump the bandwidth of the large cluster Mark was refering
to on an absolute performance basis. The question for the buyer is, "Can
I shrink my code's foot print inside the cheap-cluster node's ever increasing
cache and if so, do I have more time/talent on my hands or money." If the
answer is more money, then buy the "vector cluster".
There are a couple less dominant points like SSI, administrative costs,
percent utilization, and sustaining the capacity to design custom processor
/system in the US ... where the custom engineered vector system may also
have an edge.
More information about the Beowulf