🚀 go-pugleaf

  [Dr. David Kahaner is a numerical analyst on sabbatical to the 
   Office of Naval Research-Asia (ONR Asia) in Tokyo from NIST.  The 
   following is the professional opinion of David Kahaner and in no 
   way has the blessing of the US Government or any agency of it.  All 
   information is dated and of limited life time.  This disclaimer should 
   be noted on ANY attribution.]

  [Copies of previous reports written by Kahaner can be obtained using
   anonymous FTP from host cs.arizona.edu, directory japan/kahaner.reports.]

To: Distribution
From: 
 David K. Kahaner
 US Office of Naval Research Asia
 (From outside US):  23-17, 7-chome, Roppongi, Minato-ku, Tokyo 106 Japan
 (From within  US):  Unit 45002, APO AP 96337-0007
  Tel: +81 3 3401-8924, Fax: +81 3 3403-9670
  Email: kahaner@cs.titech.ac.jp
Re: Center for Comp Physics at Tsukuba & 300GFLOPS MPP
19 August 1992
This file is named "ccp.92"

ABSTRACT. First description of new Center for Computational Physics at
Tsukuba University, with 300GFLOPS parallel massively computer.

In an earlier report, "sacad.92, 1 July 1992, it was noted that a new
computational physics center was being formed at Tsukuba University.
The current report provides more details.

The University of Tsukuba has many years of experience designing and
using advanced computing equipment for physics calculations. I have
reported on numerous occasions about the QCDPAX and related parallel
computers. QCDPAX was both designed and built at Tsukuba specifically to
perform the computations associated with quantum chromodynamics. Its
architecture is that of a two dimensional torus mesh, one of the first
of this kind, and certainly the earliest practical mesh machine in
Japan. Tsukuba scientists built special floating point controllers and
also utilized existing chips. Various versions of PAX have been built
over a decade, and QCDPAX is the fifth of the PAX series. For parts of
QCD computations more than 14 GFLOPS has been reported, making it one of
the fastest QCD machines in the world. (In the spring of 1990, 480
processors (PU) were installed achieving a peak speed of 14GFLOPS.
However, most simulations are performed with 432 PU's to reserve some
for replacing defective units.  In this case the peak speed is about
12.5 GFLOPS.)

Other related projects are at Columbia University (NY) where a 16GFLOPS
machine is running, at IBM (NY) with an 11GFLOPS machine, at the Fermi
Lab (Illinois) with a 6GFLOPS machine, and an Italian project with a
1GFLOPS machine.  In addition, a new project with hopes of producing a
100GFLOPS QCD machine is in progress in Italy and new projects for
producing even more than that are planned in US in collaboration with
about 10 institutions including Columbia.

The PAX mesh architecture has proven to be useful for its intended
computations, but current thinking is that a somewhat more flexible
architecture is necessary for a wider range of physics applications. 

One of the leaders of the PAX project was

       Prof Yoichi Iwasaki
       Director Center for Computational Physics
       Dept of Physics
       Tsukuba University
       Tsukuba, Japan
        Tel: +81 298-53-4289, Fax: +81 298-53-4492
        Email: IWASAKI@LEPTON.PH.TSUKUBA.AC.JP

Professor Iwasaki is now the Director of a new project, titled Center
for Computational Physics (CCP), that began this summer at Tsukuba.
Recently, I met with Iwasaki and reviewed the literature associated with
CCP.

Actually there are two essentially different funding programs. CCP is a
ten year project funded by the Japanese Ministry of Education, Mombusho,
to bring together researchers to work in three specific areas of physics
where computation is a key tool.
   
   Elementary particle physics
   Condensed matter physics
   Astrophysics
together with
   Parallel computer engineering

Missing from this list is fluid dynamics, which has been the bread and
butter of large-scale computing facilities for many years. But Tsukuba
physicists have highly developed expertise in the areas above,
especially the first, built upon their years of experiences with PAX,
and they intend to concentrate on these topics. CCP will have ten full
time Professors/Lecturers, including four new positions.  About twenty
well known computational physicists (not all from Tsukuba) are also
listed as part of the project, but most will not participate full time.
These include T.Hoshino (Tsukuba) who initiated the PAX project and
Y.Oyanagi (Tokyo), who contributed heavily to many of the PAX machines.
Iwasaki explained that CCP would have about ten students working with
them, but their level (initially) will not be high.  In addition, there
are two new positions for guest (or visiting) scientists from other
Japanese institutions. Iwasaki explained that there are mechanisms for
foreign researchers to participate in the project, for example through
the JSPS (Japanese Society for the Promotion of Science).  CCP has just
become operational, and there is, as yet, no additional space or
building set aside for it. Iwasaki tells me that is one of his highest
priorities. 

Physicists working on CCP projects will need computational resources.
At the moment they have access to the QCDPAX machine, and also
supercomputer facilities at the KEK (High Energy Physics) lab.  Fujitsu
is placing five of its AP1000 parallel systems at Japanese universities,
but not Tsukuba, so access to high performance computing is not as good
there as it should be.  Additional power is needed too, and this is to
be provided by a new machine, which is being separately funded as the
major project administered by CCP.  Drawing on the QCDPAX experience,
Iwasaki and his colleagues have decided that a "practical" massively
parallel computer could be built with a few thousand nodes, each having
several hundred MFLOPS peak computing capability, i.e., a total
computing power of several hundred GFLOPS. Such a computer should have a
few tens of GBYTEs of memory and perhaps ten times as much disk storage.
It would differ from PAX not only in having a more flexible
architectural design, but also by paying more attention to high
performance input/output requirements, better peripherals, etc.

Because the PAX project is considered very successful, it was natural
for Mombusho to try and build upon that effort, and it was decided to
seek possible collaboration on a machine for related computing. Iwasaki
explained to me that there were discussions not only with the major
Japanese electronics companies, but also with some in the US. In the
end, the only company that elected to collaborate was Hitachi. Fujitsu
and NEC already have active parallel computing projects, AP1000, and
Cenju respectively.  In addition, Fujitsu has publicly announced an even
newer parallel machine with performance on the order of 300GFLOPS, so
they obviously have a plan in mind for development. On the other hand
Hitachi, the largest of the Japanese "big three" has been virtually
absent from the parallel computer field and clearly needed a vehicle. It
seems likely to me that this could be the company's serious entry into
the semi-commercial market. This will depend on timing and how
successfully the development proceeds.

Mombusho has funded this part of the project (i.e., the new machine) at
about 1.5Billion Yen (about $12M US) for five years. This is a very
large project by Mombusho standards, and in fact it was one of only two
large Mombusho projects this year.  (My opinion (DKK) is that it might
be difficult to purchase a commercial machine of this performance level
within the given budget--hence the plan to design and build one--even
though buying one, if it could be found, might have allowed more time
for doing physics. I assume that Hitachi's activities, including their
R&D are internally funded; thus the actual project budget will be
greater than that allocated by Mombusho.  The total budget for the CCP
project--non hardware--is is not yet fixed and will be requested on a
year-to-year basis, but in my opinion (DKK) it is likely to be
comparable to the hardware part.) Hitachi and CCP scientists are working
on the architectural design now, and it is to be finalized this year, so
that a running system will be available by March 1996. That is likely to
be an extension of the mesh concept that has been explored by PAX.  But
CCP scientists have already announced that their target is a 300GFLOPS
machine with an unspecified number of processors.

Iwasaki explained to me that there will be active collaboration between
CCP and Hitachi.  CCP scientists (physicists and computer scientists)
are now making a rough plan for the basic architecture of the MMP in
hopes of having a design that will allow for a high sustained speed on
application programs. Hitachi staff are also involved in these
discussions.  Hitachi will then provide the hardware, and basic system
software, such as software for send and receive, e.g., message passing.
CCP will work on the application software.  At the moment Hitachi's work
is at three of their labs, but some consolidation is likely to occur.
Iwasaki admitted that most physicists simply want to use standard
languages and are reluctant to immerse themselves in low level language
programming. It is possible that other groups at the University could
make use of the new machine, but this is not yet established.

Although the available details on this project are still quite sketchy,
there is nothing unreasonable about their plans. 300GFLOPS by the mid
1990s should be achievable, especially as the machine is not really
designed to be for fully general purpose use. It follows a trend that I
have seen before here in Japan, of focusing on specific applications to
drive the development of new computers rather than on designing a
general purpose system and then looking for users.  The key CCP people
are physicists and computer scientists, both having experience in
parallel computing.  Thus they have very specific ideas of the problems
that they want to solve.  Their track record with PAX is very good, and
by obtaining assistance from a major computer manufacturer, they are
likely to be successful in getting the machine they want, on time.  The
missing element is the software base for parallel processing, which is
not highly developed at Hitachi yet. It is likely that the company is
viewing this as an experiment and also as a way of enlisting the
assistance of a number of scientists who are already experienced with
parallel computing.

----------------------------END OF REPORT------------------------------