ARSC T3D Users' Newsletter 64, December 8, 1995

Another T3E Announcement

To go along with the announcement by CRI of the T3E, and the Wall Street Journal's T3E article, both carried in last week's newsletter, here is what appeared in the Minneapolis Star Tribune, December 4th, 1995:


  > The new Cray T3E
  > 
  > Company may get boost it needs from cheaper, faster supercomputer
  > 
  > By Steve Alexander, Star Tribune Staff Writer
  > 
  > Cray Research today will introduce what it says is the world's first super-
  > computer capable of performing 1.2 trillion calculations per second. The 
  > move caused Wall Street to take notice after a year of bad financial news 
  > from the company.
  > 
  > Cray said the computer can do in one second what a person with a handheld
  > calculator could do in 150,000 years.
  > 
  > Mardi Larson, a Cray spokeswoman, said that while other companies are work-
  > ing on machines with similar speeds, none will be available in the range of 
  > model sizes and prices that Cray plans to offer.
  > 
  > In anticipation of today's announcements, Cray's stock closed up 87.5 cents 
  > Monday at $23.125, with trading volume four times the 10-day average.
  > 
  > Cray officials confirmed Monday that they will introduce the T3E model 
  > today, which they said is the world's first supercomputer that can be scaled
  > upward - expanded with additional processors - from eight to 2,048 computer-
  > chip processors without a loss of performance because of internal computing
  > bottlenecks.
  > 
  > Cray also slashed the prices of an entry level price machine for its pre-
  > decessor supercomputer, the T3D introduced in 1993.
  > 
  > The new T3E also can operate as a stand-alone supercomputer, whereas it pre-
  > decessor had to be hooked to a traditional Cray supercomputer. Cray said that
  > will make the new machine appealing to corporations that aren't already Cray
  > customers.
  > 
  > Good news in a bad year
  > 
  > The announcement is a ray of sunlight in what otherwise has been a gloomy 
  > financial year for Cray. In October, the company reported a $13.5 million 
  > third-quarter loss and acknowledged that it would lose money on operations 
  > this year in addition to $145 million in pre-tax write-offs taken so far in
  > 1995. Cray also has sold the first of several facilities it has on the block
  > and has said it is planning further employment cutbacks and restructuring 
  > charges in the fourth quarter.
  > 
  > The T3E is Cray's second-generation machine to use parallel processing 
  > architecture, in which computer problems are broken into hundreds or thousands
  > of pieces that are simultaneously solved by off-the-shelf microprocessors
  > chips operating in parallel. The approach is much different than Cray's
  > traditional vectorprocessing architecture, in which up to dozens of propri-
  > etary Cray processors solve problems at slower speeds. But parallel processing 
  > is the trend in supercomputing, and in 1993 Cray announced it would join that
  > trend.
  > 
  > No more bottlenecks
  > 
  > Bill Minto, Cray senior director of product marketing, said that with the T3E
  > the company solved the bottleneck problems usually encountered when more 
  > processors are added. It did so by making software and input-output systems
  > more expandable, he said.
  > 
  > For scientific problems that are easily divided into many small pieces, such
  > as processing data in oil field exploration, the T3E can be scaled up to
  > thousands of processors without loss of efficiency, Minto said.
  > 
  > Cray said the T3E has a "theoretical peak" performance of 1.2 trillion cal-
  > culations per second, although customers aren't likely to quite reach that on 
  > practical scientific problems.
  > 
  > Larson said that while other firms have claimed the ability to build scalable
  > supercomputers with up to 2,000 processors, Cray was the only company that 
  > demonstrated even a 1,000-processor machine on NASA's supercomputer bench-
  > mark tests.
  > 
  > Cray said it has $92 million in advance orders for the new supercomputer, but
  > it made no predictions about how important those revenues will be in the 
  > company's projected return to profitability in 1996.
  > 
  > "We have not made any projections for 1996 except to be profitable, and this
  > will be a contributor to that," said Brad Allen, Cray's director of investor
  > relations. " In addition to contributing in 1996, the T3E also firmly esta-
  > blishes that Cray is committed to the high end of the supercomputer market."
  > 
  > Scientific use is shrinking
  > 
  > Gary Smaby,, an analyst with the Minneapolis-based Smaby Group, said Cray's
  > T3E announcement is an important signal to the firm's traditional scientific
  > and engineering customers that it is not abandoning them, even though their
  > demand for high-end supercomputers is shrinking and Cray needs to appeal to
  > growing commercial markets to achieve a financial turnaround.
  > 
  > The scientific and engineering market is shrinking largely as a results of 
  > reduced federal government expenditures for supercomputers with the end of
  > the Cold War.
  > 
  > "There had been concern among some in that [scientific and engineering] group 
  > that Cray was shifting its focus to commercial markets to the exclusion of 
  > its traditional customer base," Smaby said. "Cray's growth clearly will have
  > to come from other markets, but Cray can't afford to leave scientific and
  > engineering market behind ... This market is the foundation upon which they
  > have to build the [financial] comeback story, although the T3E in itself isn't
  > going to be the key."
  > 
  > Cray also is taking a risk in introducing the powerful new supercomputer, 
  > which Smaby said could steal sales away from some of Cray's traditional vector
  > supercomputer models.
  > 
  > "Some of the T3E systems they are going to sell could in fact cannibalize their
  > vector supercomputer business. But that was business they could have lost 
  > anyway from the vector product line, which was vulnerable to competition," Smaby
  > said.
  > 
  > Asked about that issue, Minto said the new and old product lines are comple-
  > mentary rather competitive.
  > 
  > "We see the customer selecting the most appropriate system depending on the 
  > workload. We're giving them more choice," Minto said.

Details on the T3E Processor

With the announcement of the T3E, we all are interested in the details of the DEC Alpha 21164 processor. One of the best sources of information is the DEC home page at http://www.dec.com . Below is a sample of information from that home page. At that site there is also a very nice postscript collection of overhead slides in the file, alpha-21164.ps, explaining the caching strategy on the Alpha 21164.

  > Digital To Ship Samples Of Latest Alpha Chip, 
  > World's Highest-performance Microprocessor
  > 
  > MAYNARD, Mass., November 13, 1995 -- Digital Equipment Corporation announced
  > today that it will begin shipping samples this month of the world's fastest 
  > and highest-performance microprocessor, the Alpha 21164, manufactured in a 
  > 0.35-micron process. 
  > 
  > Described at the Microprocessor Forum held in October in San Francisco, the 
  > new chip will achieve the highest SPEC92 and SPEC95 ratings of any micro-
  > processor either shipping or announced. At its design center, the new chip 
  > will deliver performance of more than 500 SPECint92 and 750 SPECfp92, and 
  > more than 11 SPECint95 and 17 SPECfp95. 
  > 
  > Digital's Alpha microprocessors have held the industry lead for performance 
  > ever since the introduction of the first Alpha chip in February 1992. The 
  > 0.50-micron Alpha 21164 is currently the highest-performance microprocessor 
  > in volume shipment; the 0.35-micron Alpha 21164 chip will keep Digital and 
  > Alpha in the performance lead through the next round of competitive product
  > introductions. Digital expects to achieve 1000 SPECint92 in 1997 with its 
  > next-generation Alpha microprocessor, which will follow the Alpha 21164. 
  > Following are SPEC performance figures for the latest products in the indus-
  > try's major microprocessor architectures, due for delivery in 1996.  Figures 
  > are based on vendor estimates. 
  > 
  >            Alpha PA8000  PowerPC 604e  UltraSPARC  R10000  Pentium Pro
  >            21164         166MHz        200MHz      200MHz  200MHz
  > 
  > SPECint92  >500   >360   225           322         300     366
  > SPECfp92   >750   >500   NA            462         600     283
  > 
  > SPECint95  >11    8.6    6.0           NA          7.4     8.09
  > SPECfp95   >17    15     5.0           NA          15      6.70
  > 
  > Pricing and volume availability of the 0.35-micron Alpha 21164 microprocessor
  > will be announced later. 
  > 
  > Digital Semiconductor, a Digital Equipment Corporation business headquartered
  > in Hudson, Massachusetts, designs, manufactures and markets industry-leading
  > semiconductor products including Alpha microprocessors and PCI chips for 
  > networking, bridging, and graphics/multimedia, as well as low-power StrongARM
  > microprocessors under license from Advanced RISC Machines Ltd. Digital 
  > Semiconductor operates design centers in Hudson; Palo Alto, California; 
  > Austin, Texas; and Jerusalem, Israel. A new, $450-million fabrication 
  > facility in Hudson will begin revenue production in 1996. Mitsubishi Electric
  > Company is a second source for Alpha microprocessors. 
  > 
  > Digital Equipment Corporation is the world's leader in open client/server 
  > solutions from personal computing to integrated worldwide information 
  > systems. Digital's scalable Alpha platforms, storage, networking, software 
  > and services, together with industry-focused solutions from business 
  > partners, help organizations compete and win in today's global marketplace. 
  > 

The T3D Reading Workstation Files

In past newsletters (#58 and #59) we described how to read and write files between the T3D and the Y-MP. In this newsletter, we look at reading and writing files between the T3D and a workstation. This extends our table of examples:

  Example      Type of    Source(writing)           Target(reading)
  (newsletter)  File

  1 (58)   Direct Access  T3D(t3d1.f)               Y-MP(ymp1.f), using asnunit
  2 (58)     "      "     T3D(t3d1.f)               Y-MP(ymp2.f), using IEG2CRAY
  3 (58)     "      "     T3D(t3d2.f),cray format   Y-MP(ymp3.f)
  4 (59)     "      "     Y-MP(ymp4.f)              Y-MP(t3d4.f), using asunit
  5 (59)     "      "     Y-MP(ymp5.f),ieee 64bits  Y-MP(t3d5.f)
  6 (59)     "      "     Y-MP(ymp6.f),ieee 32bits  Y-MP(t3d6.f)
  7 (64)   Unformatted    workstation(SGI)          T3D           
  8 (65)   Direct Access  workstation(SGI)          T3D

Unformatted Files

Although it may be possible, in particular instances, to read and write unformatted files between different machines, there should be no hope of that from what the Fortran 77 standard specifies about unformatted files. The Fortran 77 standard says that the format is 'processor dependent' and each vendor can define the format as whatever makes sense to them. But the f77 implementation of unformatted I/O on UNIX machines seems to be a 'de facto' standard and CRI has made use of it.

Example 7

unformatted files

Lets look at a simple case:


  c write unformatted file
          integer*8 ia( 10 )
          integer*8 ib( 20 )
          do 10 i = 1, 10
             ia( i ) = i
    10    continue
          do 20 i = 1, 20
             ib( i ) = i
    20    continue
          write( 10 ) ia
          write( 10 ) ib
          end
Examining the file produced on the SGI workstation, fort.10, with the unix tool od (octal dump) gives us:

  0000000 000000 000120 000000 000000 000000 000001 000000 000000
  0000020 000000 000002 000000 000000 000000 000003 000000 000000
  0000040 000000 000004 000000 000000 000000 000005 000000 000000
  0000060 000000 000006 000000 000000 000000 000007 000000 000000
  0000100 000000 000010 000000 000000 000000 000011 000000 000000
  0000120 000000 000012 000000 000120 000000 000240 000000 000000
  0000140 000000 000001 000000 000000 000000 000002 000000 000000
  0000160 000000 000003 000000 000000 000000 000004 000000 000000
  0000200 000000 000005 000000 000000 000000 000006 000000 000000
  0000220 000000 000007 000000 000000 000000 000010 000000 000000
  0000240 000000 000011 000000 000000 000000 000012 000000 000000
  0000260 000000 000013 000000 000000 000000 000014 000000 000000
  0000300 000000 000015 000000 000000 000000 000016 000000 000000
  0000320 000000 000017 000000 000000 000000 000020 000000 000000
  0000340 000000 000021 000000 000000 000000 000022 000000 000000
  0000360 000000 000023 000000 000000 000000 000024 000000 000240
  0000400
It looks like the format for the records is something like:

  32 bit word - length of data in bits 31-3
  data
  32 bit word - length of data in bits 31-3
  32 bit word - length of data in bits 31-3
  data
  32 bit word - length of data in bits 31-3
The same program compiled and executed on the T3D, produced a file, fort.10, that (when examined with od on the Y-MP) looks like:

  0000000000000 0000000000000000000012 0000000000000000000001
  0000000000020 0000000000000000000002 0000000000000000000003
  0000000000040 0000000000000000000004 0000000000000000000005
  0000000000060 0000000000000000000006 0000000000000000000007
  0000000000100 0000000000000000000010 0000000000000000000011
  0000000000120 0000000000000000000012 1000000000000000000024
  0000000000140 0000000000000000000001 0000000000000000000002
  0000000000160 0000000000000000000003 0000000000000000000004
  0000000000200 0000000000000000000005 0000000000000000000006
  0000000000220 0000000000000000000007 0000000000000000000010
  0000000000240 0000000000000000000011 0000000000000000000012
  0000000000260 0000000000000000000013 0000000000000000000014
  0000000000300 0000000000000000000015 0000000000000000000016
  0000000000320 0000000000000000000017 0000000000000000000020
  0000000000340 0000000000000000000021 0000000000000000000022
  0000000000360 0000000000000000000023 0000000000000000000024
  0000000000400 1000000000000000000000 1600000000000000000000
  0000000000420 1700000000000000000000 0000000000000000000000
  0000000000440 0000000000000000000000 0000000000000000000000
  *
  0000000010000
(where * represents at least 2 zero words in the file). For cf77 and f90, for T3D or Y-MP, the octal dump for this small case is the same. The general format is something like:

  64 bit word - length of data
  data
  64 bit word - length of data(with a one in the 63rd bit position)
  data
  3 64 bits words of termination
  padding of zero words out to a multiple of 512 words
For this simple case, knowing the order of writes on the workstation, it might be possible to extract the numbers on the Cray side with a user program, but not without a lot of effort and a great potential for error. And this could only be done treating the workstation unformatted file as a byte stream in C or a single record in Fortran. The best answer is that unformatted files do NOT transfer between different machines without special conversion routines.

CRI does provide these conversion routines implicitly on the read statement when the read statement is informed of the need for data conversion with an assign subroutine call. For example, the above file, written on the SGI workstation can be read on the T3D with:


        integer*8 ia( 10 ), ib( 20 )
        call assign( "assign -Ff77 -Nieee fort.10" )
        read( 10 ) ia
        read( 10 ) ib
        do 10 i = 1, 10
           print *, ia( i ), ib( i )
     10 continue
        end
When compiled with f90 for the T3D (it does not work with cf77 because cf77 treats integer*8 as integer*4). The example below mixes 32 and 64 bit reals and integers:

   On the SGI workstation f77:              On the T3D with f90:


        integer*8 ia( 10 )                   integer*8 ia( 10 )
        integer*4 ib( 20 )                   integer*4 ib( 20 )
        real*8    c( 10 )                    real*8    c( 10 )
        real*4    d( 20 )                    real*4    d( 20 )
        do 10 i = 1, 10                      call assign( "assign -Ff77
           ia( i ) = i                  +                 -Nieee fort.11" )
           c( i  ) = ia( i )                 read( 11 ) ia
  10        continue                         read( 11 ) ib
        do 20 i = 1, 10                      read( 11 ) c
           ib( i ) = 10+i                    read( 11 ) d
           d( i ) = ib( i )                  do 10 i = 1, 10
  20        continue                            print *,ia(i),ib(i),c(i),d(i)
        do 30 i = 1, 10                10    continue
        print *,ia(i),ib(i),c(i),d(i)        end
  30        continue
        write( 11 ) ia
        write( 11 ) ib
        write( 11 ) c
        write( 11 ) d
        end

   1  11         1.000000000000000  11.0000           1, 11,  1.,  11.
   2  12         2.000000000000000  12.0000           2, 12,  2.,  12.
   3  13         3.000000000000000  13.0000           3, 13,  3.,  13.
   4  14         4.000000000000000  14.0000           4, 14,  4.,  14.
   5  15         5.000000000000000  15.0000           5, 15,  5.,  15.
   6  16         6.000000000000000  16.0000           6, 16,  6.,  16.
   7  17         7.000000000000000  17.0000           7, 17,  7.,  17.
   8  18         8.000000000000000  18.0000           8, 18,  8.,  18.
   9  19         9.000000000000000  19.0000           9, 19,  9.,  19.
  10  20         10.00000000000000  20.0000           10, 20,  10.,  20.
SGI's f77 implementation of unformatted I/O seems to be handled correctly by the CRI conversion routines invoked by the -Ff77 flag on the assign call. Hopefully this is the case on all workstation implementations of unformatted I/O.
Current Editors:
Ed Kornkven ARSC HPC Specialist ph: 907-450-8669
Kate Hedstrom ARSC Oceanographic Specialist ph: 907-450-8678
Arctic Region Supercomputing Center
University of Alaska Fairbanks
PO Box 756020
Fairbanks AK 99775-6020
E-mail Subscriptions: Archives:
    Back issues of the ASCII e-mail edition of the ARSC T3D/T3E/HPC Users' Newsletter are available by request. Please contact the editors.
Back to Top