SUG trip report

[Steve Grandi]

The annual meeting of the Sun Users Group took place in Anaheim from
December 6 to 8. 

Announced at the meeting were two new systems: the SPARCserver 1 and the
SPARCserver 490.  As you might expect, the SPARCserver 1 is a SPARCstation
1 without a monitor packaged with several disks.  For a list price of
$18,900, one can get a SPARCserver 1 with 8 Mb, two 327 Mb disks and a 150
Mb cartridge tape. 

The interesting machine is the SPARCserver 490.  The 490 is the cabinet of
a SPARCserver 390 with a new cpu/memory system.  The buzz word was
"balanced:"  I/O capacity commensurate with cpu capacity.  The 490 uses
the same Cypress/TI SPARC/FP chip set as used in the 4/300 series but
running at 33 MHz instead of 25 MHz.  The memory system features ECC error
correction instead of the parity error detection used in the 4/300 series.
The memory bus (64 bit) runs synchronously at 120 Mb/sec (fast enough to
handle full I/O throughput plus full memory traffic even assuming 100%
cache misses).  With 1 Mb RAM chips, the 4/390 supports up to 160 MB;
multiply the capacity by 4 when 4 Mb chips are available (soon).  The
4/490 does have a P4 bus for a video adapter. 

The 4/490 also features a 128 KB write-back cache and a 4 KB I/O cache.  A
similar I/O cache was first used by Sun on the 3/400 cpu: it buffers DMA
traffic from the Ethernet and disks and provides word instead of byte
transfers of such data to memory.  The 4/390's MMU supports 4 GB per
process and 64 contexts in hardware (as compared to 16 for the 4/300) and
features hardware flush.  The VME bus runs at 22 MB/sec and the claim is
that VME throughput is not saturated even by 4 IPI disk controllers each
running 8 drives and two 1/2-inch tape drives. 

Sun's IPI-2 disk controller is their own design featuring a 68020 (20 MHz)
and 1 MB dynamically segmented read-ahead cache.  The ISP-80 controller
will support 6 MB/sec drives, but the drives they are selling with the 390
and 490 are 3 MB/sec CDC drives.  The controller can run as a VME master
at 30 MB/sec.  IPI disk cabling is much simpler than SMD cabling.  Sun's
benchmarks compared Unix sequential reads (through the file system)
comparing a 4/330 with one SCSI controller and 2 disks, a 4/370 with 4
SMD-4 controllers supporting one disk each and a 4/490 with 4 ISP-80 IPI
controllers supporting one disk each.  Relative performance for these
configurations are 1.5, 1.9 and 5.4.  The Sun guys said disk striping is
not seen as a win on their high-end IPI systems given projected upgrades
in IPI performance. 

Sun now officially sells and supports 8 mm Exabyte tape drives on 390 and
490 systems.  Sun folks claimed that these drives are fully qualified
(requiring microcode changes in the drives) and supported by a Sun driver.
Unlike the case of the SMD-4 controller card where the microcode changes
are reserved to Sun alone, Exabyte is free to incorporate the fixes for
one and all.  Sun's driver will initially be only available on the 390 and
490 feature tapes; but will be available to all with the 4.1 distribution.

The 4/490 racks up the following numbers on the various benchmark scales:
22 Dhrystone MIPS, 3.8 double precision Linpack MFLOPS, 16.8 SPECmarks and
47 TP1 tps.  Other benchmarks of interest mentioned in passing are 11.3
SPECmarks for the 4/300, 8.3 for the 4/60 and 8.1 for the 4/260. 

The SPARCserver 490 has a list price of $99,900 for a system with 32 MB
and 1 Gb of disk.  The increment over a 390 is around $25,000.  Sun claims
this is lower than the MIPS M/2000, about 1/2 of the price of a DECsystem
5810 and 1/4 the price of a VAX 6430.  There are upgrade paths for all
rackmounted Sun servers (3/180, 3/280, 4/280, 4/390).  Also, Sun has
lowered the price for additional IPI disks for the 390 and 490 (and a good
thing, too, the previous incremental cost for this disk -- a CDC
SABRE-1230 with an IPI interface -- was outrageous even by Sun's
standards!)

Sun also announced DeskSet 1.0, a suite of Open Look applications which
are apparently unbundled from OpenWindows but are similarly available for
only a media charge.  DeskSet is "based on Open Look and on SunView" and
features a "drag and drop" desktop metaphor.  The programs included in
DeskSet are File Manager ("a better way to manage your Unix files"),
Calendar Manager ("instant appointment and resource scheduling"), Mail
Tool, Print Tool, Tape Tool, Text Editor, Calculator, Icon Editor,
Snapshot Tool, Performance Meters, Clock and Binder ("a developer's tool
for defining work methods").  The DeskSet bundled with the first release
of OpenWindows lacks several of these tools. 

Sun executives were much less forthcoming about product plans than they
were at last year's SUG meeting.  All they did was hint: multi-processor
servers are coming, multi-bus servers are coming, the pizza box, the 330
and 370 pedestals and the 390/490 rack will be the shape of any new
products to come.  Sun will not leave the market to the low priced SPARC
clones expected in mid 1990; expect Sun to have both much cheaper and much
faster desktop SPARCstations than the current SS-1.  It is unclear whether
Sun is talking about one (both cheaper AND faster) or two products. 

Rob Gingell and Pat Harding gave their annual "state of Sun-OS"
presentation.  SunOS 4.1 is in Beta test and will be released when it is
qualified; probably in the spring.  4.1 is the same for all "real" Suns
(i.e., there is no 4.1c release) but there will be 4 distributions (Sun-3,
3x, 4 and 4c).  "There are no plans for a 4.1 release for the 386i" (but a
4.0.3 was mentioned).  Thought is being given to only two distributions
(Sun-3 and Sun-4) for future releases; but what the folks REALLY want to
do is bundle all releases on one CD-ROM. 

4.1 complies with POSIX 1003.1, FIPS 151-1, XPG2 and most of XPG3 (through
/usr/xpg2bin), and SVID issue 2 -- including RFS and mandatory file and
record locking (not guaranteed for other than local files!) -- and most
SVID 89 interfaces (through /usr/5bin).  The C compiler has been ANSIfied
enough to meet POSIX 1003.1; but no function prototypes.  4.1 supports
"8-bit clean" utilities to go along with the 8-bit clean kernel of 4.0.
The international features of the type-4 keyboard are supported. 

4.1 supports memory locking in the memory management facility.  Pages may
be locked through mlock(addr,len) and address spaces may be locked with
mlockall(flag).  SV compatible locking is available with plock(region).
These three calls are restricted to the super-user.  General applications
may influence the system's page replacement policy through madvise (addr,
len, advice) where the advice can be normal, sequential, random, or
will/won't be seeking. 

The system can now utilize a /tmp file system built from swap space which
has no "forced-write" semantics.  I/O to disk occurs only if memory demand
requires it; otherwise files exist only in main memory.  On reboot, these
files are GONE!  SunOS utilities have been doctored to put files that
should survive reboots into /usr/tmp instead of /tmp; users should be
similarly careful! 

Asynchronous I/O is included in 4.1 through variants of read and write
that are conceptually similar to process handling.  aioreas and aiowrite
are like fork, aiowait is like wait and aiocancel is like kill.  This I/O
is truly asynchronous and not "not-blocking".  Asynchronous I/O is a
kernel configuration option and available only to dynamically linked
programs (this is the wave of the future: in this case it preserves the
implementor's freedom to pull the 4.1 kernel based implementation of
asynchronous I/O out of the kernel in later releases). 

4.1 includes the promised programmer's interface to dynamic linking,
loadable modules (which aren't integrated into the usual autoconfigure
process, unfortunately), HoneyDanBer uucp, new AWK, an ISO 9660 (CD-ROM
High Sierra) file system, compressed kernel crash dumps (thus the
requirement that swap space must be greater than physical memory is
removed), up to 256 file descriptors (watch your select calls again!), up
to 256 pseudo-terminals, more flexible configuration options (no software
limit on number of Ethernet interfaces!), adaptive NFS retransmission
(automatic setting of timeo, rsize and wsize based on round-trip times),
mount table enhancements (providing enough information that stating a
mount point of a dead server will not hang the process), Berkeley "Fat
Fast File System," much improved interaction between YP and named,
decreased kernel residency requirements (may actually decrease behavior of
large memory systems in favor of 4 MB machines), some reworked algorithms
to minimize thrashing (such as the way the "window damage signal" is
propagated to processes), a changed default optimization level for the
680x0 C compiler (to -O2) and much improved context-switch performance on
3/200 and 4/200 systems.  4.1 will include all the pieces necessary to
reconstruct the shared system libraries and construct libraries that
bypass YP to go directly to named. 

4.1 requires a full install (but installation speed-ups have been made and
the Installing SunOS manual has been extensively revised).  The directory
structure has been once more revised to include the concept of release
levels for 4.1 and subsequent versions (e.g.,
/export/exec/sun3.sunos.4.1).  There are no longer any distinction between
domestic and international SunOS; encryption now is included as an add-on
tape.  4.0[.x] sources, binaries and objects are OK under 4.1 excluding
"invasive and many libkvm programs."  As usual, one cannot go backwards --
do NOT build on 4.1 for execution on previous releases.  Also, the "Fat
Fast File System" file system changes mean that a file system, once
converted to 4.1, cannot go back.  Drivers (or anything referencing the
user or proc structures) must be recompiled (probably no conversion).
Sun-2 systems are no longer supported. 

4.1 has been subjected to increased regression and automated testing
including external validation suites.  Interactive performance on 4 MB
diskless machines (with configured kernels, of course) is about halfway
between 3.5 and 4.0; the Sun folks are declaring victory.  For diskful 4 MB
machines and machines with more than 4 MB, performance is approximately
that obtained with 4.0. 

4.1 is touted as being the "bridge to SVr4."  Sun's SunOS/SVr4 (not the
final name!) will be the next major release of SunOS after 4.1 and is at
least a year away.  Sun has to add some features such as async I/O and
lightweight processes back into the SVr4 base and include compatibility
and migration support.  Remember, this release will default to the SV
"look and feel."  SunOS 4.x binaries WILL run (assuming one uses shared
libraries and doesn't depend on the layout of mem, kmem or a.out).  Expect
a 10% performance hit.  Sources for 4.x applications can be recompiled
unchanged on SunOS/SVr4 through a "compatibility feature" and will run
with perhaps a 2% performance hit.  A set of lint-like tolls will help
application writers port their 4.x applications to the new release. 

A new compiler release will take place early next year.  f77 v1.3 and
related compilers will feature a new commandline option, -fast, which
invokes a set of "favorite" optimizations assuming one is compiling and
executing on the same machine.  Good stuff!  Apparently, Sun will soon
start distributing (for not too many $) an "up to date" C compiler that
tracks the rest of the languages.  The usual system C compiler must be
frozen far earlier than other compilers to provide a stable base for OS
development.  Thus, for customers who want the latest version of the C
compiler (bugs fixed and optimizations improved), Sun will provide it just
like Sun provides Fortran and Pascal.  With the new compiler release, all
language pieces will occupy a new hierarchy under /usr/lang. 

Keith Bierman gave a wonderful talk on compilers and how to tune your code
for fun and profit.  This talk and the handout that accompanied it were
worth the price of the trip to the meeting. 

For the first time in the 4 years I've been going to SUG meetings , Bill
Joy did not give a talk; he was missed.  Sun VPs Andy Bechtolsheim and
Eric Schmidt both talked about how successful the SPARCstation 1 has been
("the single bestselling workstation in history"): 40,000 units have been
shipped so far and the rate of shipment has risen to 10,000 units per
month.  According to Sun SPARC now has 2/3 of the RISC marketplace: "SPARC
has won."  Sun's own business has changed as well: whereas a year ago, 80%
of Sun's sales were CISC systems (Sun-3 and 386i) and 20% SPARC, these
fractions are now reversed.  The lack of demand for the 68030 line Sun
announced in April was a surprise.  "The Motorola line must be considered
an endangered species." 

A panel of Sun execs had to respond to some rather rude questions from the
audience.  Asked if OpenWindows would ship in our lifetimes, Sun's
response was that shipments were "ramping up" so that the answer center
wouldn't be swamped by questions.  First shipments are only in the 100s
and are preferentially going to developers.  In other words, the lead time
for delivery is still semi-infinite.  Asked about the 386i line (the lack
of commitment to a 4.1 port seemed ominous to the questioner), Scott
McNealy responded that tripling performance in 18 months with the recent
announcement of the 486i upgrade didn't seem to him like they were killing
off the product.  In fact, 386i systems are priced with a lower markup
than SPARC systems.  Asked why a Wren IV disk costs about three times more
from Sun than third party vendors, McNealy responded that Sun's margins
are the lowest in the industry.  However, he continued, he thinks Sun's
prices SHOULD be restructured to charge more for items such as computers
and software and lowered for items such as memory and disks that Sun adds
little if any value to. 

The vendor show was full of the usual suspects: database and publishing
vendors, peddlers of third party peripherals and Sun itself showing off a
4/490.  Items that caught my eye were a S-bus card from Danford Corp. with
4 serial ports plus a parallel port, 4 MB SIMMS (made out of 4 Mb chips!)
for 4/60s from several vendors (still quite expensive!) and a 9U SCSI card
from Ciprico that emulates Sun's controllers for system booting. 

The next SUG meeting and vendor show will take place in San Jose on
December 2-5, 1990.  In 1991 and beyond, SUG will host two get-togethers
per year.  The 1991 dates are june 16-19 in Atlanta and December 8-11 in
San Jose.

Steve Grandi, National Optical Astronomy Observatories, Tucson AZ, 602-325-9228
Internet: grandi at noao.edu             SPAN/HEPNET: 5355::GRANDI or NOAO::GRANDI