Path: sparky!uunet!usenix!carolyn
From: caro...@usenix.ORG (Carolyn Carr)
Newsgroups: comp.org.usenix
Subject: USENIX Summer '92 Conference Report
Keywords: USENIX Association
Message-ID: <1149@usenix.ORG>
Date: 24 Jul 92 19:21:28 GMT
Organization: Usenix Association Office, Berkeley
Lines: 530
USENIX Summer '92 Conference Report
***********************************
Best Paper Winners
Doug Moen of Sietec Open Systems Division was
given the Best Paper Award for "A Discipline of
Error Handling."
Mary Baker and Mark Sullivan, UC Berkeley
authors of "The Recover Box", won the Best
Student Paper Award.
Both papers have been published in the USENIX
Summer 1992 Technical Conference Proceedings.
Works-in-Progress Report:
Organized by Lisa A. Bloch, Sun User Group and
coordinated by Peg Schafer, BBN Systems
A Hybrid Performance Model for NFS File
Servers, David N. Williams, Ericsson Network Systems,
Richardson, TX < exu...@exu.ericsson.se>
In this session we will report on a Hybrid simula-
tion model of NFS client-server transactions.
Our current environment consists of over 400
diskless SPARC workstations supported by nine
Sune 4/490 active servers. Benchmarks combined
with trial and error were the prime methods used
in arriving at the current configuration.
A number of benchmarks exist to assist in select-
ing and tuning NFS servers, but benchmarking
has its perils and limitations. Vendor-supplied
benchmark numbers are frequently suspect, and
not every organization has the resources or skills
required to achieve accurate and meaningful
results. Even after spending extensive time
benchmarking a server, the results may not pro-
vide sufficient information on how it will work in
a specific environment.
A discrete event simulation model of the NFS
client-server relationship has been built which
provides an approximate model of existing or
proposed client-server configurations. The model
allows for flexibility in changing parameters and
does not require the investment in time, and pos-
sibly money, that comes with benchmarking.
Phonestation: Moving the telephone onto the
Virtual Desktop, Stephen Uhler, Bellcore
< s...@bellcore.com>
The PhoneStation system is a tool that allows a
Sun SPARCstation to control an ordinary tele-
phone line. It consists of: 1) a micro-controller
that interfaces a telephone line with a SPARCsta-
tion via a serial port and the audio connector, 2) a
software library for the SPARCstation that pro-
vides telephone interface control, digital signal
processing (e.g., touch-tone detection), and text
to speech conversion, and 3) a TCL based lan-
guage for writing telephone applications.
As an example, the system can be used to inte-
grate answering machine functionality into the
workstation environment. Voice messages
appear as ordinary electronic mail and are played
through the SPARCstation speaker. If mail is read
from a dumb terminal, the PhoneStation system
places a call to a user specified telephone number
and plays the voice portions of any messages.
Texas: An efficient, highly compatible persistent ob-
ject store using pointer swizzling at page fault time
Vivek Singhal, University of Texas at Austin
< sing...@cs.utexas.edu>
Texas is a persistent object store that implements
huge address spaces efficiently on standard hard-
ware. Pointer swizzling (address translation) at
page fault time converts the pointers in a page
from a long format into normal, hardware-sup-
ported virtual addresses when pages are brought
into memory. This translation is transparent to
compiled programs, allowing the use of existing
compilers with little or no modification. Modern
UNIXes such as SVR4 and OSF/1 provide the
necessary control over virtual memory with no
modifications to the operating system.
Gumby: The portable, high-performance file system
that rides on the back of your Pokey file system
Sheetal V. Kakkad, University of Texas at Austin
Gumby is a simple log-structured file system
built on a normal UNIX file system. The file sys-
tem is built inside a single UNIX file, requiring no
dependencies on underlying disk geometry, so it
is quite portable. Log structure avoids the use of
a single "home" disk block for a logical file block,
allowing any block to be written anywhere. This
optimizes the file system for use with large RAM
caches, which tend to absorb most reads and
increase the proportion of writes. We intend to
experiment with reordering read-only pages as
well, to dynamically increase locality and reduce
seeks caused by read misses.
Knowledge-Based Systems Construction in C++,
Vladimir Bacvanski, Aachen University of Tech-
nology, Germany < vladi...@rwthi3.informatik.rwth-aachen.de>
The examination of an applicability of appealing
techniques from object-oriented software engi-
neering to knowledge-based systems domain is
discussed, focusing on the promising role of C++
in this context. The entrance of expert systems
into real industrial application arena has uncov-
ered weak points of the current knowledge-based
systems technology, especially the incomprehen-
sibility, poor performance, and inability to inte-
grate with non-knowledge-based systems.
The use of C++ for building technical expert sys-
tems should provide one possible framework for
overcoming the current deficiencies. The code of
a knowledge representation language is trans-
lated into C++, bringing the possibility to use
knowledge-based techniques while remaining in
the well known environment, so that developers
do not have to abandon all their skills and move
to expensive and incompatible specialized artifi-
cial intelligence workstations. Moreover, the
combination of multiple paradigms (object-ori-
ented, procedural, and the rule-based one) in the
C++ framework produces a synergetic result.
A new multi-paradigm system architecture is
examined together with mechanisms which
diminish the impedance mismatch between
object-oriented knowledge and non-knowledge-
based systems, providing interchangeability of
objects which follow different paradigms. The
object-oriented paradigm is used not only to
model the applications, but the system's internal
components as well. The correspondence
between different constructs from the object-ori-
ented and knowledge-based systems will be
investigated, showing that it is possible and prof-
itable to model knowledge-based systems with a
set of C++ classes.
Development of an event based debugger with source
level capabilities, J. G. Posthuma, J. Scholten, J. G. Wijnstra;
U. Twente, Netherlands < posth...@cs.utwente.nl>
Finding a bug in an application is time consum-
ing and expensive. For parallel applications,
debugging is even harder. The behavior of paral-
lel applications can only be understood by look-
ing at them with great abstraction. Only specified
events of the system should be presented to the
user. But such events only give a hint where a bug
could be. After this hint, the user has to look with
greater detail. He should be able to specify both
events of higher abstraction (for example com-
munication) and source level debug events.
Events are the basis of the debugger. An event is
generated each time an important point is
reached in the execution. These points can be
specified by the programmer. An event will often
be used to indicate a place where interaction
between two processes takes place, since interac-
tion is an important aspect of parallel applica-
tions. The events of all processes are merged into
one event stream. This stream can be used
directly or stored in a database for later use. For a
long running or multi-process application, the
event stream can be quite voluminous. That is
why a number of tools are provided to make the
event stream data more manageable. The pro-
grammer has the possibility of reducing the com-
plexity by specifying filters, which remove those
events in which the user is not interested.
Another important tool is the behavior recog-
nizer. A recognizer matches behavior as specified
by the user in terms of events against the event
stream generated by the application. It is not only
possible to specify the expected behavior, but also
the behavior that is not allowed. Recognizers can
be used in a number of ways. First of all they can
be used to trace down the places where the spec-
ified behavior is violated. Secondly, recognizers
are useful to summarize the behavior by replac-
ing a number of events from the stream with one
new high level event. This allows the program-
mer to analyze the system at different levels of
abstraction. A third possibility is to use recogniz-
ers to specify interesting points in execution
where some action must be performed, like a
request for process status. This last option is only
possible for run time debugging. The recognizers
are also useful for analyzing events in the data-
base.
The debugger is a mixture of an event based
debugger, source level debugge, and a behavior
recognition system. The event based part of the
debugger concentrates on distributed aspects of
the application and can be used during or after
execution. It is not always possible to debug an
application when using the event based ap-
proach. That is why the normal source level
debugger will be integrated in the design. In the
future, other features may be added, in order to
make it a complete debugger for distributed
applications.
MACH on a Physically Secure Crypto Coprocessor
Elaine Palmer, IBM Research Division, Yorktown
Heights, New York < epal...@watson.ibm.com>
I will present an overview of a secure crypto-
graphic workstation coprocessor. The prototype
hardware, named `Citadel', was completed in
1991 at IBM's T.J. Watson Research Center. It can
perform DES encryption and decryption of data
at high speeds.
We envision its use in three possible application
environments:
1. It can be embedded in a communications con-
troller (wire or fiber) to encrypt or decrypt some
or all network traffic.
2. It can be embedded in a disk controller to
encrypt or decrypt selected files or an entire disk.
3. It can be installed as a secure, general purpose
crypto coprocessor on a microchannel bus. In this
environment it can be used to encrypt or decrypt
passwords, authenticate users and requests for
resources, encrypt or decrypt transactions, or
process sensitive data from the main host proces-
sor.
In all three applications, encryption and decryp-
tion does not degrade the throughput of its host
device. The host for our prototype is an IBM
PS/2.
The hardware and kernel software are designed
to operate in a physically secure package. If the
package detects tampering, it responds by eras-
ing its encryption keys and other secure memory.
The system software is a small, multitasking
microkernel, Mach 3.0, from Carnegie Mellon
University. I will discuss the advantages and the
disadvantages we've encountered in using Mach
for this project.
BOF Report
by Rich Salz, Open Software Foundation
< r...@osf.org>
A good BOF can be one of the best things to do at
a USENIX Conference - it can keep you going all
night.
BOF stands for Birds of a Feather, as in "flock
together." It's an informal gathering of people
who are interested in a particular area. Many
BOFs are scheduled before the conference and
announced in the conference schedule. Others are
"scheduled" on site, and announced by posting a
notice on the general message board. BOFs are
not part of the standard conference track. They
are generally held after-hours and anyone can
attend. Many of them ``adjourn to the bar,'' where
the discussions can go on for hours.
This Summer Conference's BOF topics included:
a discussion of Standards (POSIX et al), Distrib-
uted Systems Administration, Gays in comput-
ing, EFF, FSF, Usenet, NNTP, UUNET, Ultrix, Alpha,
BSD4.4, BSDI, and Obfuscated C. I think one of the
best things to do is to go to an area that's new to
you. It's a great way to get practical knowledge in
an informal setting, and a good way to meet
experts in the field.
Unfortunately, I didn't do that this year: I went to
the Usenet-related BOFs because I "had to," and
others because I wanted to get a status report. So,
while there were no doubt lots of good things
happening at FSF, EFF, Distributed Systems, and
so on, I can't tell you about them.
Tuesday: News Software & USENET
This BOF was run by Henry Spencer and Geoff
Collyer. It started with an update on C News. The
Performance Release (including much work done
by Geoff for UUNET) is out, and the Clean-Up
Release won't be out for ``quite a while.'' This
next release will have a revised source tree, and
(to the cheers of the crowd) most of the build
work will be done directly in the Makefiles.
News volume is still doubling yearly, and the
growth in newsgroups is (apparently) causing
problems for some newsreaders, most noticeably
when sorting the active file or deleting a news-
group. Newsreader writers, beware: the net is
growing faster than you think!
The volume and newsreaders then led into a dis-
cussion of threads programs like trn. Geoff is
thinking about looking into the issue of threads
databases, saying ``mthreads must go.'' On the
other hand, we did get to hear the only nice thing
Geoff has ever said about NFS: ``for reading
news, NFS is pretty good.'' There was also talk
about changing the news filesystem format, to
which Geoff replied ``fix namei.''
Bruce Jones, from the School of Communications
at UCSD, is doing his doctorate on the growth of
Usenet. He has Henry's old tapes from the start of
Usenet and is trying to gauge the interest in get-
ting a CD-ROM of, say, the first A News ``Car for
Sale'' ad. If you're interested, send email to
< bjo...@ucsd.edu>.
Stan Barber spoke about the next release of NNTP.
The client and server code has been split into
pieces and the client code is in beta-test. It's
already been ported to some PCs. The new server
should go into alpha-test in July. If you have some
new feature or bug fixes, let Stan know. In partic-
ular, if you can help make it work well with C
News he'd like to hear from you. Stan can be
reached as < n...@tmc.edu>.
Wednesday Night BOFs
On Wednesday I attended the standard Usenet
and hackers BOF track: UUNET, Obfuscated C,
BSDI, and 4.4BSD. Even though each was only an
hour long, this was a long night.
Unfortunately, I missed most of the UUNET BOF. I
wandered in during a discussion of Alternet
(UUNET's commercial IP network, no traffic
restrictions). People are interested in low-cost
methods like dial-up IP service. Rick Adams
mentioned a bit about how the FBI is a customer.
People concerned about the FBI reading netnews
should make a reality check: the FBI wants to
catch serial killers, they couldn't care less about
obnoxious netnews postings!
UUNET has also written another version of UUCP.
BSDI has licensed it, and all UUNET customers
will probably be able to get it, too. The most inter-
esting thing about the UUNET UUCP is that you
can replace the front-end configuration files so
that it looks like whatever version of UUCP you
want it to. Only BSD is supported, but HDB is an
obvious next choice.
Every year Landon Noll asks the people of
Usenet to send him the most twisted C code they
can write, and in the spring and summer he and
his group evaluate the results and pick the best
(or is it the worst?) they can find. No program
could be more than 1,536 bytes of non-white-
space, and no ``cc'' line could be more than 256
bytes. Lots of whitespace was allowed this year,
which made most of the programs a little less fun
to look at. For the first time, there were more non-
US winners that those from the United States.
Every year, this is one of the best BOFs: it's very
technical, in a weird sort of way, and it's very
funny.
I also detected a decided ``tools'' bent to this
year's winners. It would have made a nice con-
trast to the FSF BOF. While GNU software does lots
of nice things, nobody will ever say it's small. At
the Obfuscated C BOF, however, we got to see a
chess program (written by Vern Paxson, the
author of flex) that reportedly held its own
against GNU Chess. There was also a make-like
program that had some novel features. Both of
these listings could fit on a single page!
The full results will be posted to the net (in com-
p.lang.c, misc.misc, and other places) in a month
or two. Landon also warned people that he and
Larry Wall are working on an obfuscated Perl
contest, which many in the crowd thought was
kind of redundant.
Berkeley Systems Design, Inc., (BSDI) is a new
start-up that is selling BSD operating systems for
the 386-family of machines. It's a small company,
still struggling to meet their weekly payroll. For
about a thousand dollars, you get the full source
code to BSD, X, NFS, and other tools - and binaries
to run it on your IBM PC or clone. This was the
most overtly commercial BOF I attended: Rob
Kolstad is an entertaining speaker, but it was
clearly a vendor presentation. It gave information
people clearly wanted to hear, however: the room
was packed. The part I found most interesting
was that USL (the branch of AT&T in charge of
Unix) is suing BSDI. While you can never be sure
when lawyers are involved, it would seem that
they are taking exception to the claim that the
Berkeley "Net II" release, upon which BSDI's
product is based, is unencumbered. I'm guessing
that BSDI was picked because they are the first
commercial venture that hasn't bought some sort
of license from AT&T. For more information on
BSDI, contact < i...@bsdi.com>.
The last BOF of the evening was the 4.4BSD BOF,
led by Kirk McKusick and Keith Bostic of Com-
puter Systems Research Group. The schedule
said that this would include a report on the
release schedule for 4.4BSD. This was very
unusual as the CSRG folks from Berkeley have
never previously announced their release sched-
ule. Anyhow, 4.4 will be available in two formats:
4.4 and 4.4 "light." The former will require an
AT&T license; the latter will contain only the
freely-redistributable source code. This will be
more complete than earlier free releases, but will
still need some work on the kernel. Both the alpha
and the final release will be available in both for-
mats.
4.4 will have lots of filesystem features: 64 bit file-
sizes (using the longlong datatype), NFS ``leases''
that make NFS more efficient and robust, stack-
able filesystems (similar to what David Rosenthal
discussed at Baltimore; the BSD work comes from
UCLA and the Ficus project), /dev/fd, the log-
based filesystem (from Sprite), and so on. It will
also make uid and gid be 32 bits, further changing
the stat structure. These changes will all be in the
alpha release because they involve changes in the
system interface. The final release will have new
TCP/IP work from Van Jacobson, the Berkeley
streams package, and probably a new virtual
memory system (from Mach). It will also contain
as many documentation updates and bug fixes as
possible. Sun has donated their shared library
architecture, and that may also be a part of 4.4. I
can't read my notes at this point, but I think the
supported architectures include the Sparc,
HP9000, Tahoe, and others.
The bad news is that once 4.4 is solidly out the
door, the CSRG Group is shutting down. They
explained that it is hard to get more funding, the
University is using BSD less, it is too big for the
current group to develop, and that the past year
has not been fun: too much politics and name-
calling. It's probably safe to say that the worksta-
tion industry would not have happened without
BSD, and that many of us would be doomed to be
filling out RPG II forms in dimly-lit cubicles.
Thanks, guys!
Thursday BOFs
Thursday night is always a questionable night for
BOFs because things are always scattershot after
the USENIX reception. This didn't stop me from
scheduling the third Usenet-type BOF of the con-
ference, however. This one concentrated on
NNTP. The NNTP protocol is being revised by an
Internet Engineering Task Force committee. Most
of the revisions are related to supporting batching
and other facilities for low-speed links. The cur-
rent draft is available for FTP from turbo.bio.net
in ~ftp/ietf-nntp. The group is not concentrating
on facilities for news-readers. There is an unoffi-
cial group working on that; send mail to
< David_asc...@brown.edu> if you are interested in
that area.
Stan Barber gave another preview of upcoming
NNTP releases and asked for some feedback on
specific changes to the client-side inews that is
part of NNTP. This led to some discussion about
news headers. There were lots of questions for me
about INN, my Usenet/NNTP implementation. I
mentioned it at last year's BOF and presented a
paper on it this year so people were fairly curi-
ous. By the time you read this, the software
should be available, however, so I won't take up
any more space on self-aggrandizement.
One last-minute BOF that was held on Thursday
was for archive maintainers. This group is start-
ing a very exciting project to make a universal
card-catalog for software available on the net.
Many of the people involved -- Rich Morin, Ed
Vilmietti -- have lots of experience with public
archives, and it sounds like they have a good plan
of attack. For more information, contact cfcl!rich.
Well, that's it. I hope you thought this useful, and
that it spurred your interest to become a full-
fledged USENIX BOFfer.
|