Hardware Reference
In-Depth Information
dimensional cube by cloning the structure of Fig. 8-37(h) and connecting the cor-
responding nodes to form a block of four cubes. To go to six dimensions, we could
replicate the block of four cubes and interconnect the corresponding nodes, and so
on. An
n
-dimensional cube formed this way is called a
hypercube
. Many parallel
computers use this topology because the diameter grows linearly with the dimen-
sionality. Put in other words, the diameter is the base 2 logarithm of the number of
nodes, so, for example, a 10-dimensional hypercube has 1024 nodes but a diameter
of only 10, giving excellent delay properties. Note that in contrast, 1024 nodes
arranged as a 32
32 grid has a diameter of 62, more than six times worse than the
hypercube. The price paid for the smaller diameter is that the fanout and thus the
number of links (and the cost) is much larger for the hypercube. Nevertheless, the
hypercube is a common choice for high-performance systems.
Multicomputers come in all shapes and sizes, so it is hard to give a clean tax-
onomy of them. Nevertheless, two general ''styles'' stand out: the MPPs and the
clusters. We will now study each of these in turn.
×
The first category consists of the
MPP
s(
Massively Parallel Processors
),
which are huge multimillion-dollar supercomputers. These are used in science, in
engineering, and in industry for very large calculations, for handling very large
numbers of transactions per second, or for data warehousing (storing and managing
immense databases). Initially, MPPs were primarily used as scientific supercom-
puters, but now most of them are used in commercial environments. In a sense,
these machines are the successors to the mighty mainframes of the 1960s (but the
connection is tenuous, sort of like a paleontologist claiming that a flock of spar-
rows is the successor to the
Tyrannosaurus Rex
). To a large extent, the MPPs have
displaced SIMD machines, vector supercomputers, and array processors at the top
of the digital food chain.
Most of these machines use standard CPUs as their processors. Popular
choices are Intel processors, the Sun UltraSPARC, and the IBM PowerPC. What
sets the MPPs apart is their use of a very high-performance proprietary intercon-
nection network designed to move messages with low latency and at high band-
width. Both of these are important because the vast majority of all messages are
small (well under 256 bytes), but most of the total traffic is caused by large mes-
sages (more than 8 KB). MPPs also come with extensive proprietary software and
libraries.
Another point that characterizes MPPs is their enormous I/O capacity. Prob-
lems big enough to warrant using MPPs invariably have massive amounts of data
to be processed, often terabytes. These data must be distributed among many disks
and need to be moved around the machine at great speed.
Finally, another issue specific to MPPs is their attention to fault tolerance.
With thousands of CPUs, several failures per week are just inevitable. Having an
