jim@TYCHO.ARPA (James B. Houser) (07/15/85)
Hi
I received a number of responses to my question about AT&T
3BXXX processors. I regret to report that for the most part people
had negative opinions of the system. I will cover some of the general
responses first and look at the specific questions I asked. I was
also informed that there is a mailing list for people interested in
3Bs. The list is (INFO-3B@BU-CS.CSNET) with requests for addition to
(root@bu-cs.csnet).
First the general stuff. There was almost universal distaste
for the way floating point is handled, especially the performance loss
due to FP emulation rather that hardware support. This was THE most
mentioned complaint. The second most popular topic was a inadequate
network support with several people citing the lack of ARPANET style
interface hardware and software. On the issue of AT&T UNIX, feelings
were mixed. Some people thought it was real UNIX while others
disagreed sharply. Several people also mentioned a problem with the
maximum block size for TAR tapes.
A few submitters like the 3Bs. A typical comment was:
"The good news: It really is unix, the 3B2s are quite impressive in
power (they are little IBM/PC looking things that sit on your desk top
and compute like a 750.) I am very glad to be able to follow AT&T into
this, I expect they will, in the long run, make us glad they did (eg.
rumours about hooking these things up to PBXs, the phone system in
general, software products etc.) I also have a UNIX/PC (AT&T 7300) on
my desk, that is *very* impressive for the price (just thought I would
mention it.) The 3B5 will soon be providing some much needed compute
power for students here, I think it will serve them well (as long as
they don't use floating point.)"
Unfortunately much more typical were negative comments such
as:
"Let me preface this by saying that all we have are 3b2's, and I hate
them."
"My time will be more than paid for if you end up not buying any."
"Unless you really want SYS V or it's free I wouldn't recommend the
3b5. Then again you can get hardware and software maintenance from
AT&T and they actually seem pretty good at it."
"It's probably *really* a good computer. They just have the ATT-IS
people supporting it (you know them, the purveyors of that wonderful
OS, System V :-) ). They're just used to having a captive audience is
all."
"You call THAT Unix? While AT&T has the name trademarked, I think of
Unix as more a set of ideals and philosophies than a software product.
In addition, since well over 80% of the utility of Unix is in the user
programs, and since AT&T has been systematically unbundling software
(vi is part of the "editor" package; others are a line printer spooler
package, a graphics package, and a "spell" package), you don't get a
lot of Unix for your money."
And the specifics:
QUESTION
How does the 3BXXX compare to Vaxen especially the 11/780 and
11/785? Is it faster/slower, more or less reliable etc.
ANSWERS
A) A 3B20S is about like a VAX-11/780, a 3B20A twice that. A 3B5 is
about like a VAX-11/750. A 3B2 is about like a VAX-11/730.
A) I believe the 3B5 does raw computing at about the speed of a
VAX750, the same for the 3B2s. The primary advantage of the 3B5 is
that it is set up for a larger configuration (as far as I know you
could not get anywhere close to that with a 3B2 [mem, disk, ports].)
A) A 3B20S is about like a VAX-11/780, a 3B20A twice that. 3B2: a bit
more than a VAX 11/750
A) A 3B5 is about like a VAX-11/750. A 3B2 is about like a VAX-
11/730.
A) Slow. Feels like a souped-up pdp-11.
A) The only one I've done benchmarks on is the 3B2. And *that* only
for cpu bound tasks, specifically,
{double,int} i; for (i=1; i<10000; i++) ;
For this, in integers, it runs same speed as our Vax-11/750, actually
a little slower. The floating point is no comparison at all
considering that they don't have hardware floating pt. The effective
rate is something like 2 or 3 flops per second. (er...more like 2 or 3
hundred anyway)
As for supposrting users, they is a 6 or 7 user machine I think.
Especially if you were to add ample amounts of faster disk space.
The AT&T people claim the 3B20 to be faster than a 780. I haven't
looked at ours yet, except to recompile part of the news software. It
worked ok for that.
A) We have not yet pushed our 3B5. I suspect it has some advantages
over a VAX for time sharing (like intelligent tty ports.)
A) In response to your first question, here are the results of some
benchmarks that Steve Shumway here at Duke has been running. These
give a fairly good indication of the speed of the 3B2/300. We may be
getting a 3B20 sometime soon, in which case we will run another set of
benchmarks to check it out. On the reliability issue, we had quite a
few problems with the 3B2's when we first got them. The problems now
appear to have been largely due to very large size files (wtmp,
LOGFILE, etc.). When proper cleanup routines were implemented most of
these problems went away. We've had them for just under a year and
their reliability is stabilizing.
Table 1. System/User time in seconds for Buchholz benchmark.
CPU CPU I/O MIX TOTAL
VAX-11/785 0/4 1/11 0/15 1/30
VAX-11/780 0/6 2/14 0/22 2/42
VAX-11/750 0/11 4/25 1/40 5/76
ATT UNIX/PC (7300) 0/15 4/10 0/52 4/77
Masscomp 0/14 9/13 1/47 10/74
Sun-2 WS 0/15 3/27 0/54 3/96
ATT 3B2/300 0/16 10/28 1/59 11/103
PDP-11/44 0/57 22/28 4/188 26/273
IBM XT (PC/IX) 0/172 39/101 7/569 46/842
Table 2. Performance relative to VAX-11/750.
CPU CPU I/O MIX OVERALL
VAX-11/785 2.75 2.42 2.73 2.61
VAX-11/780 1.83 1.81 1.86 1.84
VAX-11/750 1.00 1.00 1.00 1.00
ATT UNIX/PC (7300) 0.73 2.07 0.79 1.00
Masscomp 0.79 1.32 0.85 0.96
Sun-2 WS 0.73 0.97 0.76 0.82
ATT 3B2/300 0.69 0.76 0.68 0.71
PDP-11/44 0.19 0.58 0.21 0.27
IBM XT (PC/IX) 0.06 0.21 0.07 0.09
Table 3. Floating point benchmark results in seconds.
CPU real user system relative
VAX-11/785 (FPA) 3.8 3.4 0.1 2.057
VAX-11/780 (FPA) 14.3 4.8 0.5 1.358
VAX-11/750 (FPA) 7.7 6.8 0.4 1.000
Masscomp 11.2 11.0 0.1 0.649
PDP-11/44 (FP11) 39.0 34.7 0.2 0.206
Sun-2 WS (Sky, -fsky) 79.5 78.9 0.3 0.091
Sun-2 WS (Sky) 86.5 85.9 0.4 0.083
IBM XT (PC/IX, 8087) 95.5 93.8 0.5 0.076
IBM XT (PC-DOS, 8087) 140.72 - - 0.051
Sun-2 WS 235.3 234.7 0.4 0.030
ATT UNIX/PC (7300) 312.9 311.2 0.1 0.023
IBM XT (PC-DOS) 4024.54 - - 0.002
ATT 3B2/300 5535.1 5424.5 99.7 0.001
Table 4. Bell Benchmarks
VAX-785 VAX-780 VAX-750 Sun-2 Masscomp 3B2/300
b1 cc 0.7/1.3 0.9/1.5 1.6/2.6 1.3/1.8 1.4/1.9 1.9/3.9
run 57.1/10.2 75.7/0.4 114.1/1.6 69.7/0.0 57.1/0.2 124.2/1.7
b2 cc 12.3/2.1 19.4/2.9 32.9/4.8 28.6/2.9 30.1/4.2 62.0/8.0
run 14.7/0.1 19.2/0.2 34.0/0.6 22.9/0.1 20.6/0.1 39.2/0.1
b3 cc 2.2/1.2 3.2/1.5 5.6/2.7 4.6/1.8 5.5/2.0 8.6/4.0
run 30.5/0.1 38.6/0.1 65.3/0.8 47.4/0.1 45.9/0.1 79.4/0.0
b4 cc 1.6/1.1 2.4/1.6 4.3/2.8 3.4/1.9 4.1/1.8 6.4/4.1
run 13.9/0.1 18.0/0.1 32.4/0.6 19.4/0.0 16.8/0.1 35.4/0.0
b5 cc 1.7/1.2 2.7/1.6 4.5/2.6 3.9/1.9 4.8/1.9 8.2/4.2
run 12.6/0.1 15.9/0.0 27.1/0.4 19.0/0.0 15.9/0.0 34.0/0.0
b6 cc 1.3/1.2 1.9/1.6 3.4/2.5 2.8/2.3 3.0/2.2 12.2/7.5
run 8.7/0.1 10.8/0.0 17.9/0.2 11.9/0.0 10.1/0.0 20.1/0.1
b7 cc 1.2/1.0 1.6/1.5 2.9/2.4 2.5/1.8 3.0/1.9 4.1/4.0
run 10.2/0.1 14.0/0.1 24.9/0.4 38.5/0.0 36.9/0.1 34.3/0.0
b8 cc 0.5/1.2 0.7/1.4 1.1/2.5 1.1/1.6 1.2/1.7 1.2/4.0
run 0.0/2.2 0.2/2.6 0.8/4.3 0.6/4.1 0.1/6.0 0.8/8.5
b9 cc 0.5/1.1 0.7/1.6 1.1/2.5 1.0/1.8 1.1/1.6 1.1/4.0
run 0.4/1.8 0.6/2.2 1.3/3.8 0.6/3.6 0.2/2.3 0.5/3.2
b10 cc 0.2/1.2 0.8/1.5 1.3/2.4 0.8/1.8 1.1/1.8 1.3/4.1
run 0.2/10.5 0.4/15.5 0.7/28.1 0.3/19.5 0.3/14.5 0.5/23.3
b11 cc 0.4/1.2 0.7/1.5 1.3/2.3 1.0/1.8 1.0/1.8 1.0/4.0
run 0.1/1.4 0.2/1.6 0.5/2.5 0.4/2.2 0.1/2.1 0.3/3.4
Table 5. Compiler speed (composite of compile times for Bell benchmarks)
Machine Time rel 750 rel 780 rating
VAX-785 36.4 2.475 1.462 0.625
VAX-780 53.2 1.694 1.000 0.606
VAX-750 90.1 1.000 0.590 0.659
Sun-2 72.4 1.244 0.735 1.000
Masscomp 79.1 1.139 0.673 0.782
3B2/300 159.8 0.564 0.333 0.523
Table 6. Machine Configurations
VAX-785 DEC VAX-11/780, 4 Mb memory, FPA, 1 456-Mb disk, 4.2 BSD UNIX
VAX-780 DEC VAX-11/780, 4 Mb memory, FPA, 1 456-Mb disk, 4.2 BSD UNIX
VAX-750 DEC VAX-11/750, 2 Mb memory, FPA, 1 456-Mb disk, 4.2 BSD UNIX
Sun-2 Sun 150 WS, 2 Mb memory, Sky FP, 1 160-Mb disk, 4.2 BSD UNIX
Masscomp Masscomp MG500 WS, 2 Mb memory, FP, 1 50-Mb disk, RTU UNIX
3B2/300 ATT 3B2/300, 2 Mb memory, 1 30-Mb disk, System V.2 UNIX
UNIX/PC ATT UNIX/PC (PC 7300), 1 Mb memory, 1 20-Mb disk, Sys. V.2 UNIX
PDP-11/44 DEC PDP-11/44, .75 Mb memory, 2 160-Mb disks, 2.8 BSD UNIX
IBM PC-XT IBM PC-XT, .64 Mb memory, 1 10-Mb disk, PC/IX
Explanations
The benchmark used to produce tables 1 and 2 is the Buchholz
performance test published in the IBM Systems Journal, no. 4, 1969, p.
309. The particular version used in this test was obtained from the
USENET. It uses 32-bit integer computations. The system and user
times from Table 1 were added and compared with the values obtained
from the Vax-750 tests to obtain the relative time values of the
second table.
Input parameters for Buchholz test:
CPU 0 0 50000
I/O 20000 5000 1
MIX 4000 700 200
The floating point benchmark is a simple program that loops
250,000 times executing a double-precision (64-bit) floating point
multiply, divide, add, and subtract on each iteration. Results given
include those for a Sun-2 Workstation with and without the Sky
floating point processor board, and with and without the -fsky
compiler option.
The Bell benchmarks are a test suite of programs supposedly
obtained originally from Bell Laboratories. The focus of each test is
given below.
Tests:
b1 recursive C function calls
b2 recursive binary tree insertion sort
b3 quick sort
b4 character i/o
b5 hash table manipulation (compiler symbol table routines?)
b6 linked list manipulation (kernel block allocation routines?)
b7 arithmetic calculations
b8 kill system call
b9 time system call
b10 access system call
b11 getuid system call
The combined system and user compile times for each machine
are used as an indication of the speed of the compiler (Table 5).
Each compiler was handicapped based on system speed (obtained from the
"OVERALL" relative value from Table 1) to produce an adjusted compile
time which was then normalized by dividing by the lowest value. This
gives a measure of the efficiency of the compiler, i.e. how well does
it take advantage of the speed of the system on which it runs.
Obviously, there are many factors which influence the validity of this
measure, such as the efficiency of the generated code, the similarity
or dissimilarity of the cpu-i/o mix of the overall buchholz value to
that of the compiler, etc. Nevertheless, the figure is included as a
point of interest.
A) The 3B20 (Simplex) is about like a VAX 11/780 as far as capacity,
thoughput, etc. The architecture is much different, but that's not
important unless you run processes bigger than 16Meg, or use software
that depends on VAXisms like byte order and such. The main
limitations are in disk drives - we have a 300 MB removable (about
like DEC RM05) and a 675MB Winchester, but that's about it. No Fuji
Eagles, no RA60s,... There is some Ethernet support (3Bnet), and
Datakit LAN. The terminal driver boards are not exciting; they were
overdesigned for the 1200 baud modems everyone used to use, but you
can't fill them up if you use a lot of 9600 baud; we run ours at 4800
for most applications. Price about the same as a VAX - $250-350K if
loaded.
The 3B20A attached processor machine is master/slave dual
processor machine, a bit faster than the VAX 11/785.
The 3B20D Duplex machine is kind of like a VAX 11/782 - two
CPUs, thoughput about 1.9 times the single processor machine. It's a
very-high-reliability machine for applications like telephone switches
where you don't want them to ever go down, at least not all the way.
The 3B2 and 3B5 architecture is totally different. They are
based on our 32000 and 32100 series of 32bit microprocessor chips
(which chip is in which machine has evolved a good bit). The 3B5 has
a 32000, and is about the performance and size of a VAX 11/750. The
disk drives are different; there are 160 and 340 MB winchesters, and a
48MB "CDC Lark" drive which is half removable. It's a real nice
office machine - low power, doesn't need A/C, doesn't need raised
floor. About $70K - $100K. The 3B2/300 is the same CPU, but a much
smaller box (desktop), limited IO architecture, smaller disks.
Originally it was limited to a 10 or 36 MB disk; there's now an
expansion box for a 36 or 72 MB disk and a 24MB streaming tape drive.
Cost for the basic model $10-15K. It has 2 RS232 ports, plus 4 IO
slots. A slot can handle an IO card (4 RS232 + 1 parallel), or a
3Bnet ethernet card, or a few other things. The original CPU wasn't
real fast, and the floating point implementation was *ABYSMAL*.
They've upped the clock speed bit, and improved the compiler by a
factor of 10 (still with no floating point hardware.) The 3B2/400 uses
the 32100 series chips, and is about twice as fast - CPU performance
is about 1mips (ie 1 VAX 780) - and there's floating point chip
available. IO and disk is the same as the 3B2/300; the case is a bit
bigger so you can put in 4 Meg of memory; the 300 is limited to 2 Meg
because of room in the box. A nice machine. The 3B15, announced for
1Q86 availability, uses the same IO and cabinets as the 3B5, with a
32100 processor and a souped-up clock; about 1.4 MIPS, and it has the
floating point hardware.
A) As for up time, it's excellent. I think it has crashed three times
since we got it up in February, and at least one of those may have
been my fault initially playing around, it just runs and runs (tho
again, we haven't had many users on it yet.)
A) Before I bought a 3B20 I'd look at clusters of uVaxII's, or 8600's
first. If I were looking at 3B2's, I'd look at uVaxII's first most
emphatically. You're more likely to get a good deal documentation
wise. Also, the 3B2's we have are anything but reliable.
A) Dennis Ritchie once was heard to say (roughly), "the 3B's are
reasonable machines, but I don't know why you'd buy one of them rather
than a VAX".
QUESTION
What kind of shape is the network support in?
ANSWERS
A) With UNIX System V Release 3, due out in a few months, there will
be very nice support for networking. At present it is no great shakes
(e.g., no TCP/IP).
A) Also, not having TCP/IP is a blow for us, the good news is that it
should be available soon so that is kind of a non-issue (there has
been an official announcement by the Wollongong Group, AT&T may
distribute directly.)
3Bnet is not the wave of the future though there were some rumours
around that what we have is 'not 3Bnet' (but it's called 3Bnet...I
dunno, I guess they mean there will be a major release in the future.)
A) [censored]
A) You don't want to know. The network support is an OPTION, and the
interface you get is so low-level it's no fun. You don't get tcp or
even udp. I think (_think_--the documentation sucks, though it's in a
cute manual) that you generate IP packets. Maybe.
A) THE WOLLONGONG GROUP
NEWS BUREAU (i guess now that UPI is in trouble...:-)
For Immediate release (05/03/85)
WOLLONGONG SIGNS AGREEMENT WITH AT&T
PALO ALTO, Calif. -- To expand communications through AT&T computers,
The Wollongong Group and AT&T have signed an agreement under which
Wollongong will provide its standard networking product for 3B
supermicro and supermini computer under UNIX System V.
"Wollongong's software products will have the required Department of
Defense standard interface services for 3B users," said David J.
Preston, director of marketing and sales for Wollongong. Among
capabilities to be provided are:
--File transfer (FTP)
--Electronic mail (SMTP)
--Virtual terminal (TELNET).
"As a result, 3B users will be able to communicate over a multitude of
networks," said Preston, including Ethernet (trademarked by Xerox
Corporation), ARPANET, MILNET, the defense Data Network, point-to-
point nets, and custom-designed networking systems.
"Delivering this advanced standard of networking software to the UNIX
System marketplace is an important step in AT&T's program to become a
significant industry supplier of high-quality, state-of-the-art
computing and communication systems," stated JoAnne Miller, product
manager of 3B Networking Software. " This product family is another
step in AT&T's overall strategy of continuing to provide and support
system capabilities with the special advantage of adherence to current
software standards," Miller continued.
A) A previous employer was bidding on a contract with AT&T. I was
told by a colleague more involved in it than I was that the Wollongong
Group was doing a IP/{TCP,UDP} port that AT&T would make available as
an official Sys5 "workbench."
A) Network? What network? You call 3Bnet a network? You must be
joking! Actually....they'll be coming out shortly (months anyway)
with a TCP/IP package for 3B's. It'll be a version of the Woologong
Group's code.
A) The only answer to your second question about network support that
I can give is - "Shakey". We really have not done too much with it
yet, but that is what our experience so far has been.
QUESTION
Are there any plans to port 4.3BSD to the 3BXXX hardware?
ANSWERS
A) I haven't heard of any - and I have a nagging suspicion that it
wouldn't work too well.
A) There is little motivation for any vendor to port 4.3BSD to a 3B.
There is some chance that such a project may be started at the
University of Texas CS Dept. but why bother.
A) The biggest gripe I have in general is that SYSV is not 4.2bsd
though: 1. we have been porting a lot of the applications that we miss
so it is getting much better 2. Release 2 is much better than Release
1 which we started with.
A) when Hell freezes over
A) I hope so.
A) Try getting *real* information from AT&T about the hardware first,
then talk about porting operating system's.
QUESTION
What kind of configurations make sense to run UNIX on?
ANSWERS
A) Unix? Unix? It would be nice having Unix on a 3b.
A) I'm not sure I can give a good answer for this. We have 5
3B2/300's each with 2 Meg main memory and 30 MB hard disk running
System5.2.
A) Any 3B configuration supports UNIX. The more disk the better.
A) the usual - lots of disks, just enough tapes for backup
A) On a 3B2, get all the memory you can get. Make sure to get one of
the later models with FP chip. Look into getting faster bigger disk
drives. 2 PORTS cards will be more than you need. Look into getting a
tape drive or your person that does backup's will HATE you!
QUESTION
What experience people outside AT&T (if there are any) have
had with maintenance and repair?
ANSWERS
A) doesn't break very often, doesn't mind hot rooms
A) Nobody here uses them enough for them to break down. We have five
of them idle all the time. The poweron/poweroff sequence is
frustrating.
A) We have one 3B2 that hasn't been up without crashing for longer
than 2 weeks. The repair people would come in every so often, replace
something then have to come back a couple of days later because that
didn't fix the problem. Once, they even replaced the ENTIRE computer
-- It still breaks down. (And, no, we haven't checked the power line
yet for glitches. Besides, they advertise that computer for
*all*environments*.)
A) As for their service organization: Well, they are just starting
up. The people all mean well and try real hard but it will take time
before experience sets in, they tend to putter around with problems
rather than just rolling up their sleeves and finishing, sometimes
coming back day after day. I must say that none of our problems
stopped our system which they probably sensed and affected this (for
example, one of our disk drives was never installed right from the
start so we were running with two of them which was ok at the time and
they eventually fixed it.)
A) The one definitely good thing about working with ATT is their
service. We haven't had any major problems, but we have had a lot of
questions, and the ATT approach is good. They provide you with a (as-
far-as-I-know) 24-hour/day toll-free number that you can call for
service and information. The system works well - the person you talk
to when you call is just an operator, and they'll ask for some site
information. Then within an hour (often less) a representative calls
you back. The advantage of this over other systems I've dealt with is
that on the call-back you are talking with someone who knows about the
problem you are having.
END of Q/A
Finally, thanks to all those who contributed information. It
has been very enlightening. I will not list names so as to protect
the guilty. This is my first digestification to the net so feel free
to flame at me if I misunderstood someone's intent. As a final note I
observe that the 3B is stated by AT&T to have a "earthquake-resistant
physical design" for the benefit of West Coast types :-).
jim@tycho
-------twb@hoqam.UUCP (BEATTIE) (07/17/85)
> Hi > jim@tycho WHY WAS THIS POSTED TO MULTIPLE GROUPS!!!! GIVE THE NET A BREAK. TOM.
guy@sun.uucp (Guy Harris) (07/22/85)
> WHY WAS THIS POSTED TO MULTIPLE GROUPS!!!! > GIVE THE NET A BREAK. It wasn't posted to *any* groups. "tycho" is, I believe, a No Such Agency site on the ARPANET, and the original message and the summary were mailed to the "info-unix" and "unix-wizards" mailing lists on the ARPANET. Mail sent to "info-unix" gets received by a pseudo-account at "brl-tgr"; mail for that account gets delivered to a program which posts the mail message to "net.unix". Mail sent to "unix-wizards" gets posted to "net.unix-wizards". The guy at "tycho" may not even know that it gets gatewayed. Unfortunately, the gatewaying software can't really do a comparison of every message mailed to "info-unix" with every message mailed to "unix-wizards" to detect duplicates and post only one copy to both groups (which, unless your news software is brain-damaged, or your feed's software is, or your feed's feed's software is, or... means you only will see it once). Guy Harris