Q: What upcoming conferences cover performance management and are you presenting at any of them?
A: In early December, there are two large conferences happening at the same time. LISA '98 in Boston is a Usenix-sponsored systems administration conference, and CMG98 in Anaheim, CA is the Computer Measurement Group's annual conference. Both run the week of December 6; I'll be at CMG98. I just returned from presenting an all-day Sun performance and tuning tutorial at SANE '98, run by the Netherlands Unix User Group (NLUUG) at Maastricht, and I'll be presenting again at the Nordic EurOpen/Usenix conference in Stockholm, Sweden, February 9 to 12, 1999.
SANE '98
The first International SANE (Systems Administration and Networking) conference was well attended, and I gave an all-day tutorial on Sun performance tuning to a room full of people from all over Europe and the US. The rest of the conference had a strong emphasis on security-related issues, and apart from meeting some Sun users, I got to know people who are working on developing FreeBSD and Linux. An impressive range of speakers presented, including Bill Cheswick, Phil Zimmerman, and Rob Kolstad, among others. Among the highlights of the conference was a paper on Internet packet delay measurements -- you can get more information on this from http://www.ripe.net/test-traffic/index.html.
CMG98
CMG started out as a mainframe-centric organization, but nowadays it covers every kind of system you would find in a high-end data center environment and specializes in performance management and capacity planning for mainframes, Unix systems, and Windows NT. This year, Sun will have a strong presence at CMG98, with an exhibitor's booth and with Sun folks presenting several papers and a keynote speech. I'm also copresenting a preconference workshop on TCP performance, which will include a tutorial introduction to TCP/IP and detail how to measure and tune it on Unix, Windows NT, and mainframes.
Greg Papadopoulos, Sun's chief technology officer, will be presenting a keynote speech,
Resource Management for Unix in the Data Center.
Paper abstracts
You can view the agenda at the CMG Web site (http://www.cmg.org), so I'll just highlight the papers slated to be presented by Sun engineers.
Managing the Unix Mainframe -- Adrian Cockcroft
For many years the mainframe has been the highest capacity system
available. In the last few years high-end, Unix-based servers have
grown extremely fast, and the highest capacity has crossed
over to the Unix server. This applies to raw CPU power, memory size,
DASD capacity, and I/O rates. These Unix systems must address
high-end availability and performance management issues, but can the
Unix-based performance management tools cope?
CPU Time Measurement Errors -- Adrian Cockcroft
The most common method of measuring CPU time in Unix is to sample
the state of each CPU at each clock interrupt and accumulate
global and per-process counters. The hope is that sampling theory
will cause the average values to be accurate over time.
Unfortunately, sampling theory requires an unbiased sample, and the
clock interrupt is also used to schedule process wake-ups. This bias
causes large errors to accumulate, particularly for the CPU time
used by performance monitoring daemons, which wake up for a short
time at regular intervals and can thus hide their CPU usage from the
clock interrupt. Alternative CPU measurement techniques are used to
investigate the size of this error over a range of program types and
CPU load levels.
This paper is a development of the subject I covered in my June 1998
column.
The Ubiquitous SCSI -- Brian Wong
The small computer systems interface (SCSI) group of standards has
now been with us for 15 years. In that time, it has become one of the
dominant I/O interconnects, and it's directly configured in systems
ranging from midgrade PCs to MVS mainframes. It is almost universal
in midrange systems running Windows NT and Unix and in nearly
every two-stage storage processor. Yet its familiarity and
seemingly bullet-proof system mask a complex subsystem. This paper
describes the components of SCSI and calls attention to some common
misunderstandings about this widespread technology.
Comparing MVS and Unix Workload Characteristics -- Brian Wong
Over the years, a substantial body of knowledge has built up about
the general shape of MVS workloads in the form of processor
consumption, relative I/O content, and other metrics. Unix systems
are now being deployed in environments similar to well-characterized
MVS systems. We studied over 100 production Unix installations for
usage patterns and compared them with known MVS metrics. This paper
discusses the results of that survey, which show that MVS and Unix
usage are curiously similar but also have significant differences.
Memory Usage Instrumentation and Analysis -- Richard McDougall
Unix has traditionally had good monitoring capabilities for CPU,
I/O, and network statistics, but there's little useful information
about memory utilization. This paper introduces a new set of tools
and an accurate method of accounting for each type of memory usage
in the system. The key difference is that pages shared between
libraries and processes are measured and taken into account. This
new methodology allows the administrator to view and understand how
real memory is allocated to processes, the kernel, and file buffers,
and introduces a mechanism to size the memory requirements of
applications.
Data Warehouse Performance: A Methodology for Analysis -- Jim Skeen
How much will your data warehouse queries speed up as you add
computing power? This depends on many factors, of course, such as data
warehouse size, query complexity, data characteristics, hardware
architecture, and sophistication of the DBA and the DBMS optimizer.
This paper proposes work volumeas a new method for
predicting the effect of these factors on delivered data warehouse
performance. Work volume is a function of scan volume and join
volume. To illustrate the methodology and its application, the paper uses recent
TPC-D SF1000 results.
Adrian's performance tuning presentations
I've developed a presentation that summarizes the contents of my
book. It's quite a dense technical presentation, and the main idea
is for you to generally learn what's in the book so you can consult it
for more detailed explanations. It consists of 70 slides and takes
me one to four hours to present all or part of it. I
recorded a run-through on video, and it's available as a Sun
Reseller Training CD (Sun Experts, Fourth Edition) which coordinates the
audio track, talking-head video, PowerPoint format slides, and a
text transcript. The talk can be advanced or slowed down, but it
takes about two hours at normal speed. The CDs only work on Windows
or Macintosh systems. The same CD also has presentations on
workgroup servers and Sun's Cascade Windows NT interoperability
product. Your local Sun contact should be able to get you a free copy
by ordering part number WE 216-0. I've put both my original
presentation slides with notes (in pdf form) and the PowerPoint
version slides onto the performance page at http://www.sun.com/sun-on-net/performance.html.
Wrap up
I enjoy the immediate feedback I get at conferences; it helps me
refine my explanations of complex topics, and I get many ideas
for new problems to investigate -- more than I ever have time to
work on, unfortunately. It's also nice to put a face to an e-mail
address and meet people I've corresponded with. See
you around!
We're working on an SE3.1 toolkit update. I hope we'll be ready
to put out a publicly available beta test version soon. It supports
Solaris 7 (SPARC, 64-bit SPARC9, and x86) and has several other new
features. I expect to describe it in next month's column. In the
meantime, if you develop new code in SE you should send me e-mail
and ask that it be added to the SE developers alias. We've already
made an early version of SE3.1 available to developers.
It's also time to wish you all happy holidays and a happy New Year!
Resources and Related Links
Additional resources
Other Cockcroft columns at www.sun.com
