Searching for a Solaris Workstation - revised

David G. Hough at validgh dgh
Mon Feb 14 17:19:21 PST 1994



Obtaining and configuring Solaris Unix on a system of commodity PC components
can be surprisingly complicated.  Three possible approaches include:

*    Buy a collection of components, put it together, and try to get it to
     work with Solaris.

*    Buy an integrated Windows 3.1 system and try to get it to work with
     Solaris.

*    Buy an integrated Solaris system and try to get it to work.

     I report my experiences with the latter two approaches, which sometimes
turn out much like the first approach.  But systems from AMG and Mobius would
have booted standalone but for minor mechanical glitches.   Plug and play on a
pre-existing network requires running sys-unconfig, rebooting, and answering
the same questions you'd get installing the software - but it's much faster.

     I also report -g compilation time and -O execution time results for SPEC
and Perfect codes, comparing Solaris PC's to each other and to SPARCstations.

     This report will be posted to USENET comp.unix.solaris in four parts.
This first part only is posted to other USENET newsgroups.  Comments, sugges-
tions, and requests for copies of this report in "tbl | troff -ms" source form
are welcome to dghavalidgh.com.

SUMMARY: Conclusions reached

     You can buy collections of components, and you can buy integrated  tested
systems,  and  you do have to pay extra for the integration and test.  That is
why complete integrated and tested systems may cost more than the sum  of  the
component  prices  from  the  same source.  It is worthwhile to have an expert
spend a little time instead of you spending a lot of time, unless you want  to
become  an  expert.   If you don't want to become an expert, buy from a source
that integrates and tests the hardware and software that you plan to use.

     A diskless refurbished SPARCstation ELC provides the  best  -O  execution
performance/price  for CPU-bound scientific computation among the alternatives
considered in this report, taking no account of the cost of its server.

     A commodity PC running SLS Linux 1.04 provides the  best  -g  compilation
performance/price  among  the alternatives considered in this report.  However
the SLS 1.04 Linux distribution is not yet  an  industrial-strength  operating
system, although some parts of it such as GCC and F2C work well.

     A refurbished SS2 with Weitek PowerUp provides the best combination of -O
execution  and  -g compilation.  Measured PowerUp CPU performance is mostly in
line with Weitek claims.  But wait until the  PowerUp  timing  glitch  bug  is
fixed.

     The Mobius Protege 66 provides the best performance/price  for  CPU-bound
scientific  computation  among  the Solaris PC alternatives considered in this
report.

     66DX2 PC's are better than 50DX PC's for the kinds of problems considered
in this report, but not as much better as one might suppose.

     If you purchase a PC with MSDOS or Windows, instead of one  with  Solaris
installed and working, in order to run Solaris you'll probably have to replace
the SCSI and graphics controllers, and perhaps the ethernet, sound  card,  and
CDROM  too, if those are bundled. So take that into consideration when compar-
ing prices.

     Apple Macintosh Quadra 800's running AUX 3.0.2 are  not  competitive,  on
performance  or  performance/price,  with  PC's or SPARCstations for CPU-bound
scientific computation.  Perhaps the Unix derivative for  Macintosh  PowerPC's
will do better.

     Fast SCSI disks on a slow shared network can be faster than unshared  lo-
cal ordinary SCSI disks.

     Performance anomalies abound in networked multi-user systems.  Even  when
the network is quiet and only one user is active.

SUMMARY: Floating-point performance and performance/price

     The  following  prices  are  for  the  price  comparison   configurations
described  later,  while the following performance figures are for the perfor-
mance comparison configurations described later,  which are generally  similar
but  not  identical to the corresponding price comparison configurations.  The
performance figures measure -O execution and  -g  compilation,  floating-point
performance  primarily, integer performance secondarily, "local" disk and "re-
mote" net I/O performance to a minor extent, and graphics not at all. Graphics
performance  varies  considerably  among  these systems, which should be taken
into account when  comparing  price  and  performance/price:  thus  Mobius  PC
performance/price  figures  benefit from less expensive graphics and SCSI con-
trollers than the ATI Graphics Ultra Pro 2MB and Adaptec 1742 installed on the
AMG and Gateway PC's.

                                        -O      -O exec          -g comp
Price                         -O      ~fp92      median    %/     median    %/
 $K     System              ~fp92       /$K     perf %     $K    perf %     $K

        REMOTE disks

 4.0    RFB ELC                14         3.5      24      6.0      29      7.3
 5.4    Mobius 66DX2           13         2.4      21      3.9      33      6.1
 5.6    RFB SS2                16         2.9      31      5.5      45      8.0
 5.8    Quadra 800              5.9       1.0       9      1.6      29      5.0
 6.4    AMG 50DX               11         1.7      18      2.8      36      5.6
 6.8    Gateway Linux           9.9       1.5      16      2.4      57      8.4
 7.0    RFB SS2+PowerUp        24         3.4      42      6.0      57      8.1
 7.4    Gateway 66DX2          13         1.8      22      3.0      38      5.1
 7.6    SPARCclassic           15         2.0      28      3.7      45      5.9

        LOCAL disks

 5.4    Mobius 66DX2           13         2.4      21      3.9      38      7.0
 5.6    RFB SS2                15         2.7      30      5.4      46      8.2
 5.8    Quadra 800              5.9       1.0       9      1.6      29      5.0
 6.4    AMG 50DX               11         1.7      19      3.0      40      6.2
 7.0    RFB SS2+PowerUp        24         3.4      42      6.0      58      8.3
 7.4    Gateway 66DX2          13         1.8      23      3.1      45      6.1
 7.6    SPARCclassic           15         2.0      28      3.7      47      6.2
19.7    10/41 4.1.3+2.0.1      38         1.9      76      3.9      98      5.0
19.7    10/41 2.3+2.0.1        43         2.2      79      4.0      77      3.9
19.7    10/41 2.3+3.0          42         2.1      76      3.9      90      4.6

CONTENTS

          I. Overview

          ABSTRACT with conclusions
          SUMMARY: Conclusions reached
          SUMMARY: Floating-point performance and performance/price
          CONTENTS
          INTRODUCTION
          Unix PC information sources
          Future directions
          My business card

          II. Configuration and Price

          METHODS for obtaining Solaris PC's
          REQUIREMENTS for Solaris PC's
          SPECIFICATIONS for PC price quotes
          PRICE QUOTES - MSDOS/Windows PC systems
          WARNINGS to consider with PC price quotations
          PRICE QUOTES - Solaris PC system
          PRICE comparison configurations

          III. Installation

          AMG Solaristation
          Gateway 4DX2-66E
          Mobius Protege
          SLS Linux 1.04
          Apple Quadra 800
          Weitek PowerUp
          Current Compatibility Information: Solaris for x86

          IV. Performance

          PERFORMANCE comparison configurations
          Performance Comparisons - SPEC Ratios
          Performance tests
          Performance Comparisons - Against SPARCstation 10/41
          Performance Comparisons - Head-to-Head

INTRODUCTION

     Sun has always claimed that its workstations were competitively priced in
comparison to comparable PC's.   This was a gedanken-claim until Sun announced
that its new operating system, Solaris 2, would be ported to  PC's  with  ade-
quate  power. That raised the interesting possibility that if commodity-priced
PC's were able to run workstation software, then the advantages of  RISC  Unix
workstations  would  disappear  over  time,  due to the usual economy of scale
volume arguments.  So I decided to determine for myself how  much  a  PC-based
workstation  cost  and  how  well  it performed, at least in comparison to the
SPARCstations with which I was familiar, both to satisfy my own curiosity  and
to provide grist for a report to stimulate my consulting business.

     My perspective is that of a technical SPARCstation end  user  considering
relative  economics  of  PC's.   After all, PC's running a Unix derivative can
run a windows emulator like Wabi very effectively in order to obtain the bene-
fit of mass market applications that are much cheaper than their Unix analogs,
if available at all.  Furthermore, as an  independent  end  user  working,  in
principle, in a network of cooperating peers, I appreciate certain centralized
services like such as NIS, but I expect each end user to do his  own  hardware
and  software  administration,  as I did; some of my experiences are described
below.  This report does not consider such interesting and time-consuming  re-
lated issues as installing local printing, configuring sendmail, and uucp.

     I registered for the Solaris developers conference in March  1993,  which
included  a  copy  of  Solaris  for  x86 and a C compiler, and a discount on a
SPARCclassic, which is intended to compete with PC's. So  I  had  software,  a
cheap  SPARCstation  for  comparison,  and  a  Sun  SCSI CD drive from another
SPARCstation, so all I needed was a suitable PC.   The  separate  sections  of
this document report my investigations of

II)  PC configuration and pricing

III) PC installation

IV)  PC and SPARCstation performance

Unix PC Information Sources

     Eric Raymond (esrasnark.thyrsus.com) used to post a "PC-Clone UNIX
{Hard,Soft}ware Buyer's Guide" to news.answers from time to time.     Although
it hasn't been updated recently, it still contains much useful information for
the novice.

     If you have been living in the sheltered worlds of workstations or
supercomputers, browse through a copy of Computer Shopper to see what
commodity-pricing-based open systems is all about.

     The February 1994 issue of Advanced Systems (formerly SunWorld) has an
interesting article comparing an IBM PowerPC system running AIX and a Mobius
Pentium system running Solaris.

     If you decide not to buy a system with Solaris pre-installed and tested,
you'd better buy the Solaris for x86 software before you buy any hardware and
read everything in the documentation about hardware compatibility.  Sun's
current list of Solaris-compatible PC hardware may be obtained by sending any
email to x86-hwconfigacypress.west.sun.com.  Solaris system configuration
information is said to be available by calling 310-348-6070.

     An invaluable reference for anybody moving Sun hardware around is the
Field Engineers Handbook, available from Sun Express for $500.

Future directions

     With suitable funding this study could be extended in the directions of

Other CPUs
     Pentium and DX3, AMD, Cyrix, IBM x86's;  Alpha,  MIPS,  PA-RISC,  PowerPC
     RISCs.

Other Configurations
     ELC as 4.1.3 client or with 4.1.3 local disk; as 5.3 local disk or as lo-
     cal disk configured for cachefs.

Other Unix
     The test harness requires very solid and standard sh, make, cc, and  f77,
     as  well  as  adequate implementations of many other standard Unix utili-
     ties, SCCS, and NFS.

Other compilers
     It would be interesting to test the same version of GCC and F2C uniformly
     on all these platforms.

Other languages
     Extension of the test harness to encompass C++, Fortran-90, and  Modula-3
     would  require  existence  of a substantial body of freely-distributable,
     portable source codes for realistic scientific applications.

Windows
     Windows NT and Windows 4.0 may be inherently potent enough to support the
     test  harness,  but to be usable they will require something like the MKS
     Unix  utilities  for  MSDOS,  along  with  NFS  and   X   support.    MKS
     (inquiryamks.com)  has  announced Unix utilities available for NT as well
     as OS/2, while Sun has announced that an NT-NFS product is under develop-
     ment.   Consensys, Hippix, and DataFocus are also supposed to be develop-
     ing products in this vein.  I haven't investigated yet the POSIX layer on
     Windows  NT,  but I've heard it described as minimal - perhaps a checkoff
     item like AUX - but not as well integrated with the underlying  operating
     system as AUX.

Other reports

     Other reports available in "tbl | troff -ms" form:

        SPARCstation 10/30 and 10/41 Relative Performance 93/01/29
        SPARCstation Compiler Comparison Report 92/10/25
        Suggestions for SPEC 3.0 91/06/27

My business card

     Your system or your competitor's can be  the  subject  of  this  kind  of
study.  If interested, send for my detailed business announcement.

                           David G. Hough - validgh

                        system correctness evaluation
                        system performance evaluation
                  SPARC and Solaris compatibility evaluation
                           IEEE 754 floating-point

                                            dghavalidgh.com
                  PO Box 20370         voice (408) 997-7763
                  San Jose, CA 95160     fax (408) 997-7764



More information about the Numeric-interest mailing list