Earth Simulator Project, a Major Initiative of High Performance
Super Computer Development in Japan
Keiji Tani, Earth Simulator Research and Development Center
The Science and Technology Agency of Japan has proposed a project to
promote studies for global change prediction by an integrated three-in-one
research and development approach: earth observation, basic research, and
computer simulation. It goes without saying that basic process and
observation studies for global change are very important. Most of these
basic processes, however, are tightly coupled and form a typical complex
system. A large-scale simulation in which the coupling between these basic
processes are taken into consideration is the only way for a complete
understanding of this kind of complicated phenomena As part of the
project, we are developing an ultra-fast computer named the "Earth Simulator".
The Earth Simulator has two important targets, one is the applications
to the atmospheric and oceanographic science and the other is the
applications to the solid earth science. For the first applications, high
resolution global, regional and local models will be developed and for the
second a global dynamic model to describe the entire solid earth as a
system and a simulation model of earthquake generation process etc., will
be developed.
Taking as an example of a global AGCM (atmospheric general circulation
model), here we consider the requirements for computational resources for
the Earth Simulator. Present typical global AGCM uses about a 100km mesh
in both longitudinal and latitudinal directions. The mesh size will be
reduced to 10km in the high resolution global AGCM on the Earth Simulator.
The number of layers will also be enhanced up to several to 10 times that
of the present model. According to the resolution level, the time
integration mesh must be reduced. Taking all these conditions into account,
both the CPU and main memory of the Earth Simulator must be at least 1000
times lager than those of present computers. The effective performance of
present typical computers is about 4-6 GFLOPS. Therefore, we set the
sustained performance of the Earth Simulator for a high resolution global
AGCM to be more than 5 TFLOPS.
Reviewing the trends of commercial parallel computers, we can consider
two types of parallel architectures for the Earth Simulator; one is a
distributed parallel system with cache-based microprocessors and the other
is a system with vector processors. According to the performance evaluation
for a well-known AGCM (CCM2), it is shown that the efficiency is less than
7% on cache-based parallel systems, where the efficiency is the ratio of
the sustained performance to the theoretical peak. On the other hand, an
efficiency about 30% was obtained on parallel systems with vector
processors. For this reason, we decided to employ a distributed parallel
system with vector processors.
Another key issue for a parallel system is the interconnection network.
As mentioned above, many different types of applications will run on the
Earth Simulator. Judging from the flexibility of parallelism for many
different types of applications, we employ a single-stage crossbar network
in order to make the system completely flat.
An outline of the hardware system of the Earth Simulator can be
summarized as follows:
o Architecture : MIMD-type distributed memory parallel system consisting of
computing nodes with shared memory vector type
multi-processors.
o Performance : Assuming the efficiency 12.5%, the peak performance 40
TFLOPS (the effective performance for an AGCM is more than 5 TFLOPS).
o Total number of processor nodes 640
o Number of PE's for each node 8
o Total number of PE's 5120
o Peak performance of each PE 8 GFLOPS
o Peak performance of each node 64 GFLOPS
o Main memory : 10 TB (total).
o Shared memory / node : 16 GB
o Interconnection network : Single-Stage Crossbar Network.
The conceptual and basic designs for the Earth Simulator have been
completed during the last fiscal year. The R&D for parts and packaging is
on the way and will be completed at the end of March and followed by the
detailed design. The manufacture of the system will start in the beginning
of 2000 and the system will be completed in the spring of 2002.