Cray Supercomputer FAQ


Newsgroups: comp.unix.cray,comp.sys.super,comp.answers,news.answers
Followup-To: poster
Subject: Cray Supercomputer FAQ
From: Crayfaq0220@SpikyNorman.net (Gannett)
Date: Sat, 26 Apr 2003 14:19:07 +0100
Message-ID: <http://groups.google.com/groups?selm=1fu10vc.hn0tjm12vzg34N@1cust109.tnt5.lnd9.gbr.da.uu.net>
Summary: This article is a FAQ about Cray supercomputers.
X-Newsreader: MacSOUP 2.3
X-Trace: 1051363110 news.dial.pipex.com 4865 62.188.112.109
X-Complaints-To: abuse@uk.uu.net

Archive-name: computer/system/cray/faq
Posting-Frequency: monthly
Last-modified: April 2003
Version: 1.0.9
URL: http://www.SpikyNorman.net/
Copyright: (c) 1999 "Fred Gannett"
Maintainer: Fred Gannett <CrayFaq0220@SpikyNorman.net>

                 Cray Research and Cray computers FAQ Part 3

  ------------------------------------------------------------------------
Cray Research and Cray computers FAQ Part 3

   * What's in a name ?
   * What's in a number ?
   * Where did the first ones go ?
   * Who has/had the most Cray systems ?
   * What is a T94/SSS ?
   * What went overseas ?
   * Keeping it cool
   * What's a Mega word ?
   * What is an SSD ?
   * What is SEC-DED ?
   * Binary compatibility
   * How did users shape the design of Cray machines ?
   * Why was it hard to program Cray machines ?
   * What is boundary scan ?
   * How is a T3d different from a Beowoulf/NOW/PC cluster ?
   * What is dumping ?
   * How do you start a Cray system ?
   * Instructions for starting a Cray EL system
   * Chips off the same block
   * Could you choose the colour of your Cray machine ?
   * What was Ducky Day ?
   * What were the code names rain, gust, drizzle, cyclone etc ?
   * What was the physically smallest Cray machine ?
   * What was the physically largest Cray machine ?
   * What was the computationally largest machine that Cray made ?
   * What limits prevented Cray from building even "bigger" machines ?
   * What was the Cray connection with Apple ?
   * How many people worked for Cray Research ?
   * Who ran Cray Research
   * Was a Cray supercomputer value for money ?
   * Did Cray fail? It was bought out by SGI in 1996
   * Trademark Disclaimer and copyright notice

  ------------------------------------------------------------------------


This Cray supercomputer Faq is split into sections, Part 1 describes
Cray supercomputer families, Part 2 is titled "Tales from the crypto and
other bar stories", part 3 is "FAQ kind of items", part 4 is titled
"Buying a previously owned machine" and part 5 is "Cray machine
specifications". Corrections, contributions and replacement paragraphs
to CrayFaq0220@SpikyNorman.net Please see copyright and other notes at the
end of each document. Note: Part 3 is the only part posted to
newsgroups, the latest version of this and the rest of the documents can
be located in http format just down from:

http://www.SpikyNorman.net/

What's in a name ?

Cray Research (CRI), the original founded by Seymour Cray Jr. in 1972 bought by
SGI in 1996. Was considered a seperate business unit within SGI from 1999 that was
Sold to Tera Computer in 2000 who then changed their name to Cray Inc
Cray Computer Corporation, spun off from CRI to Colorado Springs. 1989 .. 1995
Cray Labs research arm of Cri in Boulder in the early 1980s.
Cray Communications, Not related, changed name to Anite corp.
Cray Electronics ditto.
CraySoft, a venture to exploit the compiler technology and expertise of the
application software development group. This lead to the interesting fact
the each time a program was compiled on a T3e the compile script checked to
see if the machines was running Solaris!
SuperTek, original developer of the XMS, was bought in 1989 by CRI
Floating point systems, original developer of the CS6400, folded in 198?,
assets bought by CRI. Was sometimes known as Cray Research Superservers (CRS)
SSI, Supercomputer systems Inc, founded by Steve Chen chief architect of the YMP.

Crayfaq@spikynorman.net, Caryfaq@spikynorman.net, Newsposter@spikynorman.net, 
These email addresses have been discontinued and blackholed due to the huge 
quantity of spam sent to them. 
Please use the current email address CrayFaq0220@.... for communications about this
faq. 

What's in a number ?

All CPU chassis, SSD chassis and in the larger machines IOS chassis had
serial numbers but the CPU chassis number was taken as the main system
designation. From the serial number it is possible to get a general
indication of the model type as different ranges of machines were grouped
into ranges of numbers. The structure of the model designation varied but
often indicated the max. number of CPUs that could be installed, the chassis
size, and sometimes the memory size.

Cray Research System Serial numbers
        Machine type
<100    Cray 1
1nn     XMP 2 CPUs
2nn     XMP 4 CPUs
3nn     XMP 1 CPU
4nn     XMP 2 CPUs
5nn     XMP 1 CPU (SE)
6nn     Cray/ELS XMS 1 CPU
10nn    YMP 8 CPUs, Model D IOS
11nn    XMP/EA 4 CPUs
12nn    XMP/EA 2 CPUs
13nn    XMP/EA 1 CPU (SE)
14nn    YMP 2 CPUs, Model D IOS
15nn    YMP 4 CPUs, Model D IOS
16nn    YMP 2 CPUs, Model E IOS
17nn    YMP 8 CPUs, Model E IOS YMP8I
18nn    YMP 8 CPUs, Model E IOS YMP8E
19nn    YMP 4 CPUs, Model E IOS
20nn    Cray 2, 2/4 CPUs
2101    Cray 2, 8 CPUs
24nn    Cray YMP/M94 4 CPU
26nn    Cray YMP/M92 2 CPU
28nn    Cray YMP/M98 8 CPU
3nnn    SV-1
40nn    C90 16 CPUs
42nn    C92A 2 CPU
43nn    D92A 2 CPU
44nn    C94A 4 CPU
46nn    C94  4 CPU
47nn    D94  4 CPU
48nn    C98  8 CPU
49nn    D98  8 CPU
5nnn    ELs
60nn    T3d
61nn    T3d Included in a YMP2E + IOS
62nn    T3d
63nn    T3E 300 - liquid cooled
65nn    T3E 600 - air cooled
66nn    T3E 600 - air cooled
67nn    T3E 900 - liquid cooled
68nn    T3E 900 - liquid cooled
69nn    T3E 1200 - liquid cooled
70nn    T94   4 CPU
71nn    T916 16 CPU
72nn    T932 32 CPU
9nnn    J90
95nn    J90 32 CPUs

Notes:

The < 100 series can be further subdivided, among the Cray-1, Cray-1A,
Cray-1S, and Cray-1M lines. SN101 was the first XMP with 2 cpus

Eugene R. Somdahl posted in c.u.c The 1Ms were 1Ss with MOS memory. I
believe only six were built. Plus one that wound up as "scrap" running CTSS
for several years in a basement in Chippewa Falls. This was the machine on
which much of the Y-MP (and later Chen's MP project) design work was done.

XMP/EA: systems had 500 series serial numbers. Production ran from sn501 (19
Jul 87) through SN515 (15 Jan 89).

Cray-2: Later, a small number of variants occurred: 2 and 8 processors, 128
MW and 512 MW, Dynamic RAM (DRAM) and Static RAM (SRAM) don't know what SNs
Only Serial numbers Q1,Q2 and 2001..2029 built.

SN1001 Had slower clock speed than later YMP 8s.

SN1040: Originally a Model D IO system it was converted to Model E IOS for
use as a T3d host then later went to Moscow.

Cray-2 SN2000 (only a single CPU) and SN2101 (only 8 CPU) now reside in the
computer museum at Moffett field CA. There were 4 single cpu Cray2 systems
built, Q1 -> Q4. Only one was shipped to a customer, one was in Eagan and 2
were used as test systems in Chippewa.

The 1700 series was known as a Y-MP8I because it could hold 8 CPU's with 1
SSD Section and 4 IOSE clusters in a single integrated chassis. One 1700
chassis was rewired to hold three 1600 type machines. That wiremat wasn't
pretty. Served in CCN as software development machines ICE, FROST and
SUBZERO.

The 1800 series known as the Y-MP8E with similar module layout to a 1000
series but could hook up to a Model E IOS. It would typically hook up to a
700 series Model E IOS which had a chassis very similar to an 1800 that
holds 4 SSD sections and 8 IOSE clusters.

IOSE/SSDE boxes usually came in three serial number ranges 7xx, 8xx and 9xx.
The 8xx is a short chassis like the 1600 that held 1 SSDE section and 2 or 3
I/O clusters. The 900 chassis looked like a 700 chassis but was only wired
for SSDE.

A couple of these machine types (1600, 1900 and 6100) could support a JSSD
(SSDE32i/SSDE128i). That was a 32 or 128 Mword SSD that fit in a single
chassis slot.

The T3D 6100 series integrated 128 EV4's 2 Y-MP CPU's and 4 IOSE clusters
into 1 chassis, and a 6200 which I believe is like a 6100 without the
Y-MP/IOSE integration.

SV-1 may comprise of a cluster of systems resulting in the use of more than
one serial numbers.

SV-1 SN3001 to SN3007 and SN3213 to 3220 ... note that SN3217-3220 comprise
SN3501 (super-cluster) at NIH.

In 1978 5 Cray-1 systems were installed. In 1996 350 Cray J90 systems where
shipped the large part of the total of 415 J90 systems. Some J90 systems are
being converted to SV1 chasis just to keep the records complicated.

If you know sub ranges of serial numbers or special variants from the above
pattern please email the document author. Details of C3, C4 and CRS6400
numbers would also be appreciated.

The small serial number ranges indicate that the phrase "Hand built in
Chippewa Falls" was never far from the truth. With SGI winding down
development of the big Cray machines will we ever see Cray SN10,000?

Where did the first ones go ?

The early adopters of a technology are the crucial customers that can make
or break a developing technology company so it is interesting to look where
the first few Cray-1 machines were installed. In 1978 5 Cray-1 systems were
installed.

The first Cray one SN1 went to LASL in 04/76 for a six month evaluation 
after which a short-term contract was signed. In 09/77 SN1 was replaced 
with SN4, a 1000k word machine with SECDED memory protection hardware. 
SN1 Ran an OS called DEMOS written in MODEL.

SN2 was gutted and never shipped.

Looking at the Cray 1 installed base in the 1976 to 1979 time frame.

Date, Place, application
04/76, LASL Los Alamos NM, Nuclear research later replaced by SN3
07/77, NCAR CO, Atmospheric research
10/77, ECMWF UK, Weather forecasting initially SN1 later replaced in 10/78
01/78, US DOD, Defence research
07/78, US DOD, Defence research
04/78, NNFECC National magnetic fusion energy centre Livermore CA, magnetic fusion
research
09/78, United computing systems Kansas City MO, Commercial computing services
10/78, MOD UK, Defence research Used SN1 replaced 04/79
01/79, Lawrence Livermore Lab Ca, Nuclear research
04/79, Cray Research Inc., Development and bench marking

The machines above were 500k or 1000k Words. These machines were so far
ahead of their time that many of them worked at a number of sites, sometimes
being leased for short periods until the customer-ordered machine was ready.

Most non-government Cray-1 systems were front-ended by CDC CYBER 170
systems, IBM MVS or VM machines or DEC VAX systems.

Apart from the first Cray-1, which travelled the world, the first machine of
most product types typically served in the Cray computer centre (CCN) first
at Mendota Heights and later at 655 Lone Oak Road, Eagan.

The first (prototype) Y-MP/8D, sn1001, served in the Cray computer centre as
the compute platform "Mist" ( later changed name to "Gust" as "Mist" is a
rude word in German ).

The first (prototype) X-MP, sn101, remained on the floor of the 1440
building until 1989 (17 Aug 82 - 25 May 89). ... It was subsequently
scrapped.

The first Y-MP2E, sn1601, travelled to the UK and served in the Cray UK data
centre as a software development machine.

Pittsburg super computer centre (PSC) had the first customer T3D and T3E.

Cray 2 customers included NMFECC, NASA Ames, University of Minnesota, US
DOD, Harwell Laboratory UK, Aramco, University of Stuttgart in Germany and
Ecole Polytechnique Federale de Lausanne (EPFL) in Switzerland.

Quoted from From HPC wire Subject: 9037 CASH-STARVED CRAY COMPUTER CLOSES,
SEEKS CHAPTER 11 Mar. 27 CCC did make one tentative sale, to Lawrence
Livermore National Laboratory (LLNL) for support of the Department of
Energy's nationwide research program. In December 1991 when CCC was unable
to meet delivery/performance goals, the order was cancelled.
A small (4-CPU) CRAY-3 was later placed at the National Center for
Atmospheric Research (NCAR) where, after some time and effort, it became
ready for production work. This was not followed, however, by an actual
sale.

Who has/had the most Cray systems?

According to the SuperSites list in Aug 1999 The NSA runs a lot of Cray
machines.

 National Security Agency, Fort Meade, Maryland,US
       1)      Cray T3E-1200 LC1084             1300.8
       2)      Cray T3E-900 LC1328              1195.2
       3)      Cray SV1-18/576                   576
       4)      SGI Origin2000/250-864            432
       5)      Cray T3E-1200 LC284               340.8
       6)  3 * Cray T932/321024                  174
       7)      Cray T3E-750 LC220                165
       8)      Cray T94/SSS-256K                 100
      ... snip
       20) 4 * Cray C916/161024                   64

I can't possibly comment but the only other place where so many Cray
machines have been co-located is in the 655-D machine room of CCN located in
the Cray Eagan building (now Wham!net). The population of that centre
changed over time but will have had just about every sort of Cray Research
machine at one time or another. As the CCN data centre was used to house and
manage customer machines there was often more than one of each current
production system.

What is a T94/SSS ?

The SSS designation on a T90 indicates a special configuration that was
developed for particular customer. Its actual "speciality" may be seen in a
section of an ARPA document on the net. The document mentions a processor in
memory development for the Cray-3 for use in " This machine would be suited
for parallel applications that require a small amount of memory per process.
Examples of such applications are laminar flow over a wing, weather
modeling, image processing, and other large cale simulations or
bit-processing applications." It is thought (by me anyway) that this
development was updated into T90 technology.

What went overseas ?

To the eastern block.

   * An EL went to Prague.
   * 2 ELs and later a YMP-4E and two J90s went to Poland.
   * An EL and later a YMP-8E went to Moscow.
   * A C90 went to China.
   * An XMP went to India for weather forecasting.

All with the agreement of the export control authorities.

The first mainframe class Cray machine was installed in Moscow the same week
that the Moscow branch of "Planet Hollywood" opened.

International subsidiaries as of December 31, 1995. It could be reasonably
assumed that the offices were for the supply and support of Cray Research
computers in the host country. Not in the list is the UAE and the large
Aramco site that had a Cray-2 followed by a C98-DRAM machine.

International offices
Cray Research A.B.                         Sweden
Cray Research Scandinavia A/S              Norway
Cray Research (Australia) Pty. Ltd.        Australia
Cray Research B.V.                         The Netherlands
Cray Research (Canada) Inc.                Canada
Cray Research Europe Ltd.                  United Kingdom
Cray Research France S.A.                  France
Cray Research GmbH                         Germany
Cray Research Japan, Ltd.                  Japan
Cray Research (Korea) Ltd.                 Korea
Cray Research (Malaysia) Sdn. Bhd.         Malaysia
Cray Research de Mexico, S.A. de C.V.      Mexico
Cray Research OY                           Finland
Cray Research, S.A.E.                      Spain
Cray Research S.R.L.                       Italy
Cray Research (Suisse) S.A.                Switzerland
Cray Research (UK) Ltd.                    United Kingdom

Keeping it cool

The development of Cray cooling technology allowed each technology
generation to increase the circuit board density. "Someone (perhaps Gary
Smaby? I truly don't remember) once said that Cray Research was primarily a
refrigerator company."

Cray-1: Single sided boards clamped to copper plates placed in aluminium
racks that had cooling fluid in tubes.
XMP: Double side sandwich boards clamped to twin copper plates placed in
aluminium racks which had cooling fluid in tubes.
Cray-2,3,4: Immersion cooling. The CPU and memory boards sat in a bath of
electrically inert cooling fluid.
YMP, C90, T3d LC, T3e MC: Double-sided circuit boards clamped to hollow
aluminium boards in which the cooling fluid circulated.
El,J90,T3eAC,SV-1: Blown air cooling.
T90: Immersion cooling. The CPU and memory boards sat in a bath of
electrically inert cooling fluid.

So the main forms of cooling were conduction to external cooling, conduction
to internal coolant, blown air cooling and total immersion cooling. There
was some research done into sprayed coolant methods for J90 modules,
reported in an internal technical symposium paper, but this did not make it
into a publicly available product.

What's a Mega word ?

All Cray machines (except the CS6400 range) had 64 bit words. Each word
could hold 1 integer, 1 floating point numeric value or 8 characters. All
arithmetic and logical operations were done in 64 bit maths. Memory sizes
were generally designated in Mega Words. 32 MWd can be thought of as 256
Megabytes. However really the memory was 72 or 80 bits wide at the hardware
level See SECDED.

What is an SSD

Bigs gobs of memory that are attached via very high speed channels to vector
CPUs. These extra memories, often 4 to 16 times the size of central memory
could be utilised as disk IO cache, swap space, extra memory segments for
programs or even RAM based file systems. These memories were accessed on a
YMP at up to 1000 Mbytes/sec and are often used to transparently hide the IO
required for an out-of-memory solution. Sites that used SSD as root/usr file
system disk cache often saw little or no physical disk activity even when
the system was stressed. A trickle sync mechanism was employed to prevent
these vast disk caches from becoming stale.

Some later SSDs were built from J90 technology memory boards and some from
T3E boards.

What is SECDED ?

Single bit Error Correction, Double bit Error Detection. This was a hardware
scheme used in all Cray machine (except the very first C1 and CRS systems)
to allow single bit memory errors to be effectively ignored and for
detecting other memory errors to prevent memory data corruption. Each 64 bit
word had 8 other bits that were used in this advanced parity
checking/correction method. If a double bit error was detected a syndrome
byte indicated where the error had occurred. SECDED on the C90 has 64+16
bits and allowed correction at the 4-bit level.

Binary compatibility

Binary compatibility, the ability to run programs compiled on one machine on
another, extended back one generation only. You could run XMP binaries on a
YMP but not on a C90. The J90 counted like a YMP.

There were some cross compilation libraries available so that you could make
binaries for different architectures i.e. it was possible to build YMP
native codes on an XMP. Cray always recommended recompilation on the newer
system in the case of an upgrade in order to take advantage of the new
hardware features.

Execution compatibility across the range looks like :

C1 --> XMP --> XMPema --> YMP --> C90 --> T90
                   |       |       ^
                   V       V       |
                  XMS --> ELs --> J90 --> J90se --> SV1
                           *
                           *
                           T3d ==> T3e
APP --> CS64 --> CS64000
C2  .   C3  .  C4

For binary compatibility you can go across one arrow or up/down one arrow
only.

Note ==> the T3e used the same programming models as the T3d and so was
source code compatible with only minor exceptions.

Note ** An MPP emulator was available for ELs to develop MPP codes while
waiting for the T3 product line to arrive. The emulator managed about 4 MPP
nodes but used the native CRI arithmetic.

Note The C90 could run J90 binaries if compiled without the scalar cache
optimisation. The T90 comes with either Cray Floating Point CPUs or IEEE
CPUs (or both).T90 IEEE arithmetic CPUs will not support C90 native codes.

The C4 had IEEE FP and went back to lots of registers instead of local
memory. There was no binary compatibility between the C2, C3 and C4.

How did users shape the design of Cray machines ?

System owners asked for and got machine instructions to do "population
counts", leading zero and later a bit matrix multiply functional unit. BMM
was provided as an option on C90 and standard on all T90 CPUs.

Hardware based gather/scatter instructions were included in the EL and later
systems. This instruction allowed the compilers to manage array indexed data
more efficiently.

See also What is a T94/SSS ?

Why was it hard to program Cray machines ?

There has always been lots of Unix src code sloshing round on the net since
well before Linux and the admirable open src code movement was invented.
However there were a couple of things about Cray machines that made porting
codes to Cray machines tricky. We won't get into what you had to do to your
algorithm to get the best out of a Cray machine but just examine a few
things that made the conversion of codes to Cray machines interesting.

Firstly there was the word size, one rather large size fitted all, integers
and floats were represented in 64 bits, whereas most other machines used 16
or 32 bit for ints and floats. This would not cause a problem in well
written codes unless assumption were made about the range of an integer or
the relative sizes of integers to character data types.

Cray PVP machines are word addressable, the T3D and T3E are byte addressable
machines. 
Although the compilers used transparent word division to mimic byte addressable
constructs, advanced character handling was never a natural process on a
Cray machine.

When the original arithmetic units were designed for Cray PVP systems, the
IEEE floating point standard for arithmetic overflow and underflow did not
exist. The IEEE standard compromises performance in favour of ease of use
and so may not have made it in anyway. The Cray floating point format
provided a different range of numbers and level of accuracy that tripped up
some programs.

There are also other subtle considerations that have caused headaches for
many programmers porting codes to Cray machines. Known collectively as the
"Cray effect", they are the combination of algorithm scaling problems,
cyclic accumulation of errors and parallelism interdependencies that seem to
show up most times you take an apparently well behaved small program and run
it longer harder and further than possible on a conventional system.

A source adds :
    Possibly the most subtle problem in porting
    programs to UNICOS was that some very "bright"
    person put a #define for "WORD" in a header
    file required by the kernel.  The first
    thing to do in porting a program to UNICOS
    was to find if it used WORD as a #define
    or as a typedef and work around that.

    Cray machines could not support "alloca", so
    minor magic had to be applied to programs
    using "alloca."

    In the very early days, many C programs
    suffered from the "nUxi" problem, but that
    was hardly unique to Cray machines.

At various times, the following languages were available on Cray machines,
FORTRAN 66,77,90,HPF, Cray C, Standard C, C++, Pascal, Ada, Perl, but never
GNU C or Basic. There was an implementation of APL for the CRAY-1, and an
implementation of SNOBOL 4. Neither was ever "officially supported." LANL is
reported to have ported Common LISP as documented in Dick Gabriel's MIT
Press PhD thesis.

What is boundary scan ?

This is a low level logic feature that is used to initialise and test the
logic state of a board in a machine. Originally introduced on the EL it was
later adopted and used on the J90 and T90 machines.

Scan files are unique to each CPU/memory board type. On the early EL systems
boundary scan tapes with pre-built scan initialisation files for various
part numbered CPUs were available but had to be updated whenever a later
revision of board was used.

As the number of CPU module types mushroomed, Cray engineering realised the
service problems associated with scan tapes. Releasing the scb command and
the associated ASIC data files structures allowed service personnel to
rebuild scan files of a machine in the field.

If you know that your hardware is sound you can use scb command to
regenerate boundary scan files, as long as you have the /install/asic files
and scb command.

The bscan command is used to read and write the scan state off/onto a board
to progress hardware problems.

How is a T3d different from a Beowulf/NOW/PC cluster ?

This section could also have the title "The return of the Killer (network
of) micros." The battles between the proponents of the "network of
workstations" and the "it's only a supercomputer if it costs more than 1M$
per CPU" crowd rattle endlessly round the halls of comp.sys.super but we are
not going discuss the merits or otherwise of MPPs V. Vector supers. This
section is purely to describe some of the differences between modern
implementations of MPP technology. The term cluster can be considered to
refer to any collection of generic parts pushed together to make a compute
engine. e.g. Beowulf, NOW, SP2. An understanding of the applications that
you will be using in an MPP system is vital when deciding what performance
attributes are important to you.

There are four main areas of differences between a cluster and a T3e,
        . Interconnect bandwidth ( and low latency ),
        . Single system image,
        . Programming models, and
        . I/O ability.
Some of these differences may erode, as cluster technology develops,
but for now these differences stand to justify the extra 0s on the price.

Interconnect bandwidth and low latency: To make effective use of an MPP
computer or cluster the problem/program has to split the work between
multiple CPUs. Inevitably the whole task will require some communication
between the nodes working on the different parts of the answer. Sometimes
this co-ordination is only a very small part of the process, with just a
distribution of sections of the problem, at the start, and the collection of
the answers at the end, but with other types of problem close
synchronisation and large amounts of inter processor data passing is
required.

This is where understanding the speed and latency between the processing
nodes becomes important. Inter processor synchronisation speed depends on
network latency. Inter processor data passing speed depends on network
bandwidth. Total communication bandwidth depends on the interconnection
topology and sophistication of the routing algorithm.

The amount of communication required by a cartoon film render farm, for
instance, would be at the low end of the communication requirements scale
whereas a large finite element or fluid dynamic problem would be at the top
end requiring low latency interconnects and high bandwidth. Combining these
factors is the ratio of communication time to processing time. Delays in
synchronisation and bandwidth become less significant if the calculation
time between communication is long.

See The Gannett's law document for a description of a possible method of
assessing problems and cluster computer solutions.

Single system image: The T3e has a single system image, this is not to be
confused by the fact that the operating system is distributed across an
number of CPUS. One of the main benefits is that there is just one OS to
upgrade, monitor, boot and tune. Another even more important benefits is the
availability of global memory data segments, closer synchronisation between
the processing elements and shared filesystem access. The downside of a
single system image is a high degree of inter node dependence that can make
hardware problems tricky to find and fix. The close PE co-ordination makes
command and task load balancing fast and efficient.

Programming models: The hardware support of a barrier tree mechanism allows
fast multi-node synchronisation. The eureka hardware support allows problems
with wide solution branching to be co-ordinated on "first to find the
results" basis. Hardware support for atomic swap (and block memory move)
allows for fast and accurate semaphore operation between shared memory
locations. Clusters support the message passing programming models but also
seen in the T3e are the shared memory programming models popular from
symmetric multiprocessor systems.

Processor interconnect: The high level of processor and network integration
allows the T3e to perform multi-node process moves, checkpointing and
program debugging. Problems that require 1000s of process synchronisation
per short problem step can only be realistically considered on the T3e.

IO ability: Clusters can work well in the cases where the IO is limited or
centralised but many super computer applications are defined by large
amounts of computation and huge amounts of data. The T3es provision of large
shared file systems that makes sharing of huge datasets between
computational processors easy. Support for huge raid disk arrays equally
available across all CPUs in the system make shared file access routine. NFS
cross mounting between 100s of processing nodes just doesn't provide the
strength for many big problems. The emergence of multinode access
filesystems on the SP2 have served to make IO a closer race.

So to sum up a T3e is best suited to problems whose parts require very high
levels of synchronisation, co-ordination and communication. If the
requirement is for chunks of computer power loosely coordinated over long
time steps and budget is a big consideration generic cluster technology
should be considered.

What is dumping ?

Copying memory contents to disk, copying disk contents to tape or selling
computers at less than cost to stuff the opposition. CRI filed a government
investigated dumping complaint against NEC over the supply of a system to
NCAR. In a later quirk of economic fourtunes Cray Inc now sells the NEC 
machines into the Amerian market.

How do you start a large Cray system ?

This is the short version, the full version fills a very long chapter and a
whole day on the "Unicos kernel internals" course. The exact details varied
a bit from generation to generation but the principals remained the same.
Starting from a cold machine, the first thing to do would be to call the
Cray engineer to start the power conditioning MG set. Next check the
cooling circuits and WACS panel on the side of the main cabinet, to see that
the power supply rails and temperature levels are normal. Working from the
MWS the engineer would then master clear the machine to initialise the
processor and IO channel logic. After this the first code would be copied
main memory and the DEAD START command issued. The first code would usually
be diagnostics to set and check the internal logic of the CPUs and memory.
This diagnostic stage was largely replaced by the boundary scan logic in
later systems.

Finally the system would be handed over to the operators for the booting of
Unicos. The Unicos boot process went like this. Firstly all the CPUs are
halted and, starting from word 0, the kernel is copied into main memory from
the OWS or support disk, along with a parameter file describing the hardware
environment, kernel parameters, disk and IOS layout. At this point CPU 0
would deadstart and run while the rest of the CPUS idled. The kernel parsed
the parameter file and used it to locate the IOSs and root partition, both
of which it would need for single user mode. Once in single user mode the
rest of boot sequence ( triggered by init 2 ) was much like most Unixs,
spawn the service demons, initialise the networks, check and mount the rest
of the file systems, spawn the console gettys.

There were many an interesting reasons why systems would not boot but most
of them came down to hardware problems, incorrect parameter files and Unicos
bugs on new hardware. Finding what was causing a system boot failure was of
course a source of endless amusement.

Instructions for starting a Cray EL system

Check the "big red button" is up. Switch system power on at the power
breaker on rear mchine wait until system ready indicator lights up on front
panel of the machine. Press the system reset buttons under the flap on the
front of the machine. Let the console messages and diagnostics scroll up.

It should come to the BOOT> prompt. Type "load" this will scan the hardware
looking for disks and controllers according to the config.sys file on the
IOS disk. It will also run diagnostics if this is a cold start.

At this point you should have a IOS> prompt. You can now cat /bin/boot Look
for the line that has something like
lu /sys/sys.ymp /sys/config.uni

/sys/sys.ymp is the kernel the other is the text parameter file which
describes the disks/filesystems that Unicos will use. Then type "boot" to
start Unicos into single user mode. The system should now arrive at the #
Unicos prompt. Check your file systems with /etc/gencat. When this is
complete go to multi user mode with init 2

Typing ^a on the console switched you between talking to the IOS and talking
to the Unicos console unless either the scroll lock (right next to the
delete key) had been pressed or the console line had been switched into
remote support mode in which case the machine just looks hung.

A computer based learning training course for system operators was
constructed that simulated the boot sequence and other basic EL maintenance
procedures. See the annotated boot sequence in a near by document.

Chips off the same block

What was the difference between the Alpha chips used in the T3[de] and other
Alpha chips ? Basically not much except in the liquid cooled T3[de] versions
the chip was packaged upside down. Normally the heat from the Alpha
chip is dissipated through an air heat sink on the top of the chip but with
the T3e providing cooling from inside the board the packaging was flipped to
have the heat go down instead of up. DEC also provided a pin to allow CRI to
run data fetches big-endian rather than little-endian.

Heat| ++++++  chips
    V ======  PCB board
    --------  Cooling fluid in hollow copper module
    ^ ======  PCB board
Heat| ++++++  Chips

There was some consternation in 1997 when Intel bought the Alpha chip
fabrication plant, the prospect of putting "Intel Inside" on the case of the
Cray T3e filled purists with horror.

What was Ducky day ?

This was an employee fun day that happened once a year at the Eagan
buildings. The day was a holiday for all staff with organised sport and
social events. The Tee shirt design of the day was decided by a cartoon
competition. The origins of the "Ducky day" name came from a prank; a plastic duck
was found in the rather grand water pond with sculpture outside the 1440
building. The facilities manager issued a rather stern warning so the next
day the pond was filled to the brim with yellow plastic ducks.

A Revisionist view of Ducky day From Bill P.
Ducky Day began at the Mendota Heights facility (1440 Bldg.) and was moved 
to the Eagan facility with the construction of the new campus. The combination 
fountain/sculpture was christened "Octal".  It was made of wooden lattice work, 
concrete and plumbing (probably procured from a local Menards Building Center).
On the company move to the "new" Cray Research Park  from the 1440 Northland Drive 
facility,  the first ducky day at the new location was celebrated by
putting the torch to the wooden portion of the reassembled sculpture on the island 
in the middle of Cray Lake. This "Viking like" ceremony was presided over by Bob 
Ewald, the VP of Software Development at that time.


What were the code names rain, gust, drizzle, cyclone etc ?

Up to about 1994 the computers in the Cray Research, Eagan and Chippewa
Falls, machines rooms in were known by their serial numbers. For
manufacturing this was fine but for the Eagan centre it was a pain as
computers were replaced and upgraded on a reasonably regular basis. CCN in
Eagan was more interested in running compute environments so it was decided
that each type compute platform benchmarking, filestore, batch engine etc.
would have an environment name instead. This meant that "Rain" as a compute
platform started out as a YMP8e but later transparently changed into a C90
with less disruption to the end users. The exception to this was USS the
Data Migration file server that was always called USS. The 
biggest machines in Eagan were interconnected by very high speed networking,
HYPER channel in the early days, and later (@1994) 200MB/s HIPPI
connections.

Could you choose the colour of your Cray machine ?

In the early days of the company yes, there was even rumour of a cowhide
covered XMP delivered to a Houston oil company. As time went on this was
dropped as a customer option. Well almost - when there is that much money
changing hands, if enough fuss is made, the exterior panels would revisit
the paint shop. This did not apply to Els which were all black and red. Well
almost - one customer which had just upgraded a pair of YMPs (one green, one
blue) for a C90 and an EL did manage to get the EL painted a rather fetching
sky blue colour.

As for XMP/EA, sn501 and internal contact reports "We lobbied to have it
done up in denim (like the denim Jeans) & have a little red Levis tag
attached. ... Management was not amused & it never happened."

One second user customer did have a bit of a surprise when their second user
C90 arrived in a lurid deep rose/pink colour. The top of the C90, being a
convex shape happened to sit just a couple of feet under a set of strip
lights and resulted in lovely pink glow over a whole section of the machine
room.

The Bell Labs Cray (XMP) was a wallpaper'ed IC design.

What was the physically smallest Cray machine ?

The smallest, commercially available machine, was the EL92 a repackaged
version of the air cooled EL range which measured approx. 1.2m high by 0.6m
wide by 0.6m deep and could run from a normal power outlet. Whilst not a big
commercial success as it was a bit late to market, it was widely used for
trade shows, software development and as loan equipment. Available in 2 and
4 CPU versions with 512Mb memory it was truly a deskside Cray. Good write up
and pictorial in "Advanced systems" magazine July 1994.

What was the physically largest Cray machine?

The 16 CPU version of the C90 was a truly big machine standing at 2.5m tall
and CPU cab just fitting in a 4m diameter circle. Together with the power
and cooling equipment the whole system weighed in at approx. 12 (?) tonnes.
There was one (or more?) systems delivered that consisted of four
interconnected 16 CPU C90s.

What was the computationally largest machine that Cray made ?

In December 1998 the biggest machine was a 1048 CPU T3e assembled by Cray
and temporaly released to a small community of scientists. This machine was
subsequently delivered to an undisclosed customer. The design at that time would
scale up to 2176 Cpus before node addressing limits were reached.

The "Guinness World records 2000" has the Cray C90/16 as the fastest general
purpose vector-parallel computer but we all know this should be updated to
the T90/32.

What limits prevented Cray from building even "bigger" machines ?

By bigger in supercomputing terms we mean more CPU horsepower in one system. 
Looking at the T3e range of machines, we see that that the T3e was a very 
scaleable architecture. The systems can be grown by adding boards of 4 or 
8 processors during a maintenance slot. However adding boards does require 
some rewiring of the interconnect torus and takes a few hours. 

There would be various reasons why a limit is set on the upper size of a T3E. 
I am not sure which of the following played a part in the upper size limit 
of a T3e but here are the ideas.

Physical size limits: 1048 CPUs, the largest known configuration, would fill 8
system cabinets and beyond that the distance across the frame could have adverse
timing influences on the clock circuits. The clock was carried from cab to cab
using Fibre optic circuits. The physically bigger a system gets the longer it
takes to co-ordinate the parts. Which is why computers get denser by preference
rather than larger. Seymour Cray always concentrated on making smaller denser
machines rather than larger ones with more in because of this factor.

Power and cooling: not really a limit, bigger, hotter machines had been built
by Cray in the past. However beyond certain wattage very special and usually
expensive power and cooling arrangements have to be made and this will contribute
to large infrastructure and running costs.

Reliability: Because of the tightly bound interconnect of the machine, very much
tighter than nodes on a network, the failure of any CPU or board power supply
could kill the machine. If you had a CPU/power supply combination that fails once
randomly distributed in 50 years (50*52=2600weeks) / 2176cpus = 1.19 So you can
expect to see a failure just over once a week.  Building a board that fails less
than once in 50 years is hard but a system failure once a week is unacceptable.
Driving up that reliability curve for the machines was something that the
engineers battled with during the development of the system. Having "Spare" CPUs
ready to map in limited the down time associated with a CPU/cpu failure but it
still took time to locate the point of failure and map around it. A system
running weather predictions to a tight time scale could not afford to have random
2..4 hour failures once a week.  After an initial burn-in period boards do not
fail randomly but you can see that the more cpus/parts there are the less
reliable a machine will tend to be.

Operating System scaleability: The T3e is a single system image machine, even
though parts of it are distributed between processors, there is only one copy of
the operating system, process table memory map etc.  A "normal" SMP type
architecture will struggle once the number of active processes exceeds 2000 and
it has the advantage of uniform access to all its memory map. The T3E has to cope
with just as many processes with the extra strain of them all running in
parallel. The scheduling was made easier by dividing the processors into
application pools and by treating a group of processors working on a single task
as a single job but controlling the machine gets harder the more parallel
elements there are to co-ordinate. The bigger the machine the more likely it is
that you will hit a situation where much of the machine is pounding on a single
part of the machine that runs a critical resource. The cost of having that a big
machine all waiting on the completion of a single task limits the scaleability of
the system. E.g. Root filesystem inode contention, process table interactions
etc. These and other throttle points limit the scaleability of any operating
system.

Market demand: The amount of CPU horse power you can get out of one CPU was such
that most problems fitted into 50..100 CPUs before the program bogs down waiting
on I/O or system resources. If you have 40 lots of 50 CPU programs to run you
would have a more reliable solution having 2 * 1000 CPU boxes.  The demand to run
2000 CPU problems just is not large enough to cover the extra costs of building
them.


What was the Cray connection with Apple ?

Cray and Apple, seemly at opposite ends of the computer spectrum, do have
some subtle links. It was known that Seymour Cray used an Apple desktop some
of the time when designing the Cray-2. It is also known that Apple had a
sequence of Cray machines starting in March 1986 with an XMP/28 followed by
another XMP in Feb 1991. A YMP-2E arrived later in 1991 and finally an EL
from Dec 1993 to Jun 98. It is said that Apple's first XMP was bought by
Steve Jobs after he just walked into the Cray facility in Mendota Hights.

Originally purchased to help out on a computer on a chip project, the
machines eventually earned their keep running MOLDFLOW an injection plastic
modelling program ( producing some results in the form of Quicktime movies)
and later as a file server. Other applications were CFD codes for disk drive
design improvement and one source reports ".. they sometimes ran the first
XMP as a single user MacOS emulator ... They had a frame buffer and a mouse
hooked up to the IOP."

What is less known however is that the small active display panel on the T3d
was an Apple powerbook. The powerbook ran a Macromedia presentation showing
the T3e cube of cubes logo with an orbiting growing/shrinking sphere. The
display at one site was changed to alternate with a presentation plaque
display. It was rumoured that one site engineer ordered a collection of
spare bits that, over time, comprised a complete new powerbook.

According to a CCC inside source Seymour Cray and the Cray Computer 
Corporation used Macintosh desktop computers almost exclusivly for 
work on the Cray-3 and Cray-4 projects. Much of such work was just 
moving text and graphic files arround on a shared network.

The recent (Sept 1999) launch on the www.Apple.com web site of the G4
Macintosh computers displayed a YMP-8D computer on the processor page.
Whilst there was no direct reference to that particular machine there was a
requote of the Seymour quote about "using an Apple to simulate the Cray-3"
in a sidebar. ( prob this should be Cray-2 ED). The G4 is being touted as a
"Supercomputer for the desktop" and with the performance figures of a
Gigaflop/s (1 CPU) it is certainly up to at least 1992 supercomputer cpu
speed. The YMP pictured on the site would have had 0.333 Gflop/s per cpu but
was sold as sustaining 1 Gflop/s, for the whole machine, on real life
applications. It remains to be seen if the G4 can match the memory size,
memory bandwidth and IO capacity of this 8 year old Cray. Supercomputers
these days do a Teraflop/s. There is however no doubt that it will be cheaper
to buy.

The popular Macintosh telnet program developed by NCSA has an icon which
is an XMP surrounded by a network with Macs. NCSA had a Cray
accessed by Macs and thus needed to develop such a program. 
NCSA = National Centre for Supercomputer Applications.



How many people worked for Cray Research

This compiled from various SEC 10-K documents.

As of December 31, 1995, the Company had 4,225 full-time employees; 83 in
development and engineering, 1,399 in manufacturing, 763 in marketing and
sales, 915 in field service and 165 in general management and administrative
positions.

As of December 31, 1994, the Company had 4,840 full-time employees;1,172 in
development and engineering, 1,531 in manufacturing, 849 in marketing and
sales, 1,075 in field service and 213 in general management and
administrative positions.

As of December 31, 1993, the Company had 4,960 full-time employees: 1082 in
development and engineering, 1798 in manufacturing, 728 in marketing and
sales, 1157 in field service and 195 in general management and
administrative positions.

                        Cray Research Growth Table

 YearEmployees Locations USA/International New Systems Installed systems
 1972    12                 1                   0              0
 1973    24                 2                   0              0
 1974    29                 2                   0              0
 1975    42                 2                   0              0
 1976    24                 4                   1              1
 1977   199               6 / 1                 3              4
 1978   321               8 / 1                 5              8
 1979   524               10 / 3                6             13
 1980   761               13 / 4                9             22
 1981   1079              16 / 4               13             35
 1982   1352              17 / 4               15             50
 1983   1551              18 / 6               16             65
 1984   2203              20 / 8               23             84
 1985   3180             22 / 10               28             108
 1986   3999             28 / 12               35             138
 1987   4308             34 / 13               43             113
 1988   5237             35 / 18               52             220
 1989   4708             32 / 18               57             240
 1990   4857             34 / 18               52             263
 1991   5395             35 / 21               71             309
 1992   4895                                                  446
 1993   4960      6 + 153600 ft^2 Leased                      505
 1994   4840      6 + 190600 ft^2 Leased                      638
 1995   4225      6 + 211100 ft^2 Leased

Data from Cray Research via www.cbi.umn.edu.

CCC had 350 workers at the time of its closure in 1995.

Who ran Cray Research ?

CEOs of Cray Research

Who                    Dates       Where are they now ?
John Rollwagen        1976..1992   Now is a "Venture Partner" at Paul Venture
Capital in St. Paul, Minnesota
Marcello Gummucio     1992..1994   Now at BeOS ?
John F. Carlson   Jan 1993..1994
Robert Ewald      Dec 1994..1996   Now at Estamp an internet venture selling
snailmail postage over the web.
J. Phillip Samper May 1995.. ?       ?
Irene Qualters         ?

Mr. Ewald ... joined the Company (SGI) in 1996 as President of Cray Research
and a Senior Vice President of the Company. Mr. Ewald was the President and
Chief Operating Officer of Cray Research, Inc. from 1994 until 1996, when
Cray was acquired by the Company. Prior to that, he was Chief Operating
Officer of Cray's Supercomputer Operations and in 1993 served as Executive
Vice President and General Manager, Supercomputer Operations. From 1991
through 1993, Mr. Ewald was Cray's Executive Vice President, Development.
His daughter worked for part of the Craysoft organisation.

Les Davis can certainly be considered one of the spritual leaders of the
traditional Crayons.

Was a Cray supercomputer value for money ?

The utilization of most of the 16 cpu C90 systems I ever saw approached 98%
at most sites, month after month, year after year so despite the high,
typically USD 30,000,000 price customers got value for money.

Warning salesmen maths ahead, full of apples=oranges assumptions but go with
me for while...

Lets work it out over 6 years the typical life of a C90. USD 30,000,000
purchase price + 3,000,000 * 6 Per year service cost = USD 54,000,000 total.
That would be 9,000,000 per year over 365 days = USD 24,657 per day for the
system. Per CPU that would be approx. USD 1,541 per day. In that day on that
one CPU you could do 1 Gigaflop/s * 24 * 3600 * 98 % = 84,672,000,000,000
Flops in the day.

C90 Cost per 10^6 flops = 154,100 / 84,672,000 = 0.0018 cents per MFlops

Think that a workstation in 1994 would cost you approx. USD 8,000 + 800 PA
and would be obsolete in 3 years. So that is USD 10,400 total or 4366.6 PA.
Per day that's USD 12. Typically workstations are person driven, and people
work 9 to 5 - 1 hours in a day, 5 days out of 7 that's a 40/168 = 23 %
utilisation. Workstations in 1994 could manage about 5 Mflops/s on real
codes. So the workload works out at 5,000,000 * 24 * 3600 * 23% =
99,360,000,000 Flops in a day.

PC Cost per 10^6 flops = 1200 / 99,360 = 0.012 cents per MFlop.

So it is cheaper to work things out on a C90 than a workstation just as long
as you have enough work to keep it busy. We won't go into the fact that you
can run much bigger problems on the C90 or that you can get the same work
done faster as we seem to have won this point already.

Did Cray fail? It was bought out by SGI in 1996

No. Cray dominated, after inventing the super computing marketplace, for 20
years. Cray was *the* name in high performance scientific and engineering
computing. Cray did seem to lose the technology focus in the later part of
the '90s, the T90 was too expensive to build, the T3e cost too much to
develop and the DEC deal scuppered the chances of world domination in
departmental compute servers but there is no doubt that the sheer quantity
of engineering and science completed on Cray hardware changed the face of
the 20th century.

So did Cray Research fail? No. Whilst relatively few machines were ever
made, even of the more popular models, they provided an unrivalled platform
for tackling the toughest computational problems, which is of course, the
mission of Cray Research.

In March 2000 the Cray Research name and business was sold by SGI to Tera Inc, 
the inovative supercomputer maker from Seattle. Tera in have now changed their name
to Cray inc and continute to develop and service Cray computers.









Trademark Disclaimer and copyright notice

Thank you for taking time to read these important notes.

Cray, Cray Research, CRI, XMP, YMP, C90, T90, J90, T3d, T3e, Unicos, plus
other words and logos used in this document are trademarks which belong to
Silicon Graphics Inc. and others. There is nothing generic about a Cray
supercomputer.

Some of the ideas described in this document are subject to patent law or
are registered inventions. Description here does not place these ideas and
techniques in the public domain.

I wrote this document with input from a number of sources, but I get both
the credit and the blame for its content. I am happy to read your polite
correction notes and may even integrate them with the text so please
indicate if you require an acknowledgement.

Personal use of this document is free but if you wish to redistribute this
document, in whole or in part, for commercial gain, you must obtain
permission from and acknowledge the author.

  ------------------------------------------------------------------------
Things to do: Detailed machine type specs, URls and links to resources, War
stories - Time warp, Fixed on site, T3e OS internals, T3e logical to
physical, T3e go faster bits express message queues, Multi CPU Process
relocation, How to program, vectorisation, macro tasking, micro tasking,
auto tasking, Large memory, expand ccc, expand SPARC superserver,
  ------------------------------------------------------------------------
April 2002 V1.0.9

Copyright (c) 1999 by "Fred Gannett", all rights reserved.
This FAQ may be posted to any appropriate USENET newsgroup, on-line service, web
site, or BBS as long as it is posted in its entirety and includes this copyright
statement.
This FAQ may be distributed as class material on diskette or CD-ROM as long as there
is no charge (except to cover materials).
This FAQ may not be distributed for financial gain except to the author.
This FAQ may not be included in commercial collections or compilations without
express permission from the author.

Downloaded from Gannett's home page
Send corrections/additions to the FAQ Maintainer:
Crayfaq0220@SpikyNorman.net (Gannett)

Last Update May 01 2003 @ 00:59 AM