My review of O'Reilly's latest clusters book published at
HPCwire (http://www.tgc.com/hpcwire.html):
'Crazy Talk' Clutters New Cluster Book
Glen Otero, Linux Prophet
When my colleagues and I heard that O'Reilly was releasing
another
cluster book ("High Performance Linux Clusters with OSCAR, Rocks,
openMosix & MPI"), we knew it would not turn out well. One of my
colleagues even said, "It's going to be written by some guy that
doesn't know anything and [gets all excited] over clusters."
Why such a pessimistic prediction?
For one, it was uttered by the same cluster expert that O'Reilly
ignored while producing their first cluster book debacle several
years
ago. When told that their first book ("Building Linux Clusters" by
David Spector) should be scrapped and rewritten, O'Reilly ignored
their reviewers. The advice only came from the knowledgeable folks
at
VA Linux, *the* cluster company at that time. But what does VA Linux
know? It's O'Reilly, they obviously know better.
The first O'Reilly cluster book was a complete disaster. I wrote a
scathing review of it for Linux Journal in 2000. Completely void of
anything useful, the book and included software were simply not
finished. It was like reading a rough draft. Totally embarrassed,
and
suddenly void of hubris, O'Reilly apologized to its audience and
pulled the book from print.
Not satisfied to sit around pointing fingers and complaining, I told
O'Reilly I would help them with their next cluster book attempt, if
there even was one. Before long, I signed a contract to write a
clusters book for O'Reilly. But in their infinite wisdom, they
didn't
like the first few chapters that I submitted. Although I had gotten
other cluster experts to review what I had written, O'Reilly didn't
bother to get any experts to review what I was writing. They just
didn't like it, so they dismissed it out of hand. Needless to say
the
"we know better" attitude was back, and that ended the contract.
Which brings us to present day. This latest cluster book suffers
from
the same brain damaged, hubris-driven process at O'Reilly. Just like
the first book, it's written by a virtual unknown in the cluster
community (Joseph D. Sloan) and comes across as having been written
in
a vacuum.
Let's start with the book's title, "High Performance Linux Clusters
with OSCAR, Rocks, openMosix & MPI." There's nothing
high-performance
about this book because there's no discussion of using any high
performance networks like Myrinet, Infiniband, or Quadrics outside
of
four paragraphs on page 40. There are so many ill-informed sweeping
generalizations made about cluster networks on that page that I
threw
the book against the wall when I read them. For example, Quadrics
and
Infiniband are clearly established networking technologies, not
merely
"emerging," as the author believes. Sloan obviously hasn't attended
a
Supercomputing conference in the last several years. Unfortunately,
the rest of the book is rife with several inaccurate cluster
oversimplifications and incorrect definitions of terms like single
system image (SSI) and virtual machine interface (VMI). The
"beginner's guide" design of the book is no excuse for inaccuracies
and oversimplifications.
In my eyes, this book was doomed for the trash after page 8. Sloan
states that the term "Beowulf" is a politically charged term that
would be avoided in the book. That is the most ridiculous thing I
have ever heard. It's impossible to take that comment seriously,
especially since the author doesn't even take the time to properly
define a Beowulf. For these reasons alone, I can't take this book
seriously. I've thrown back my share of adult beverages with Don
Becker, and trust me when I say that the political nature of Beowulf
has never come up. Adding to the confusion, the phrase "more
traditional Beowulf-style cluster" is then used on page 63. I hope
now
you'll understand why I think this book is schizophrenic at best.
Defining a Beowulf shouldn't have been too difficult for Sloan. He
could have used a term that he introduced on page 10, "asymmetric
cluster." But I guess it's too much to ask that the Beowulf
project,
Tom Sterling and Don Becker's brainchild that started the high
performance cluster phenomenon, be properly described and defined in
a
clusters book. By the way, I've never heard the term "asymmetric
architecture" used when describing clusters. And, outside this book,
you won't either.
After page 8, it's apparent that the author has nothing original to
offer and is going to regurgitate what has already been written
about
clusters. There is absolutely no value in this because the online
documentation for all of the cluster projects covered by the author
is
far more informative than what is included in the book. For example,
while screenshots of a cluster install are included in the online
Rocks documentation, they are omitted in the book. Furthermore,
after
regurgitating much of the online Rocks documentation, the author
doesn't offer any additional helpful hints or troubleshooting
advice.
As someone who runs a company that provides and supports cluster
software based on Rocks, I can tell you that there are plenty of
pitfalls that should have been mentioned.
This underscores my major complaint with this book. There's nothing
new, nothing novel and no real help offered. Everything is just laid
out superficially in front of the reader for them to make the right
cluster decision. The book should guide the cluster decision-making
process, but it only offers a bunch of questions -- with no
substantial answers.
Sloan even admits on page 91 that there is a very detailed set of
installation instructions for OSCAR, including screen shots,
available
online. So why is this book necessary again? Oh yeah, the author is
supposed to help the reader decide if OSCAR, or any cluster toolkit
for that matter, is right for the reader. Unfortunately, no help of
any kind is offered.
The typos and omissions weren't rampant this time, but the errors I
found on pages 76, 123, 127, 130, and 136 provided nasty flashbacks
of
the first O'Reilly book. Good thing I resigned myself to do a shot
of
tequila after every typo I found. It dulled the pain this book
inflicted.
OK. "Part I -- An Introduction to Clusters" is just inaccurate and
infuriating. "Part II -- Getting Started Quickly" contains recycled
and reformatted content easily found for free online. "Part III --
Building Custom Clusters" isn't really about building custom
clusters,
but looks more closely at some software that was gleaned over in
Parts
I & II. While I don't agree with the inclusion of the parallel
virtual
file system (PVFS) and the omission of Sun Grid Engine in Part III,
I'm sure this can be chalked up to one of the tough decisions the
author had to make, like the omission of PVM and Condor from the
book.
"Part IV -- Cluster Programming" is actually a very good
introduction
to programming, debugging, and profiling MPI programs.
It's obvious that this book has no clear identity. It's like a 5th
grader's book report: a lifeless facsimile of what's been read,
totally void of originality, wisdom or topic advancement. But it's a
quick read because it uses small words.
Should I be this harsh? After all, cluster computing is a complex
subject where the answer to most questions is "it depends."
However,
I believe that O'Reilly owed us an excellent book after their first
cluster gaffe, so I'm disappointed that O'Reilly took the easy way
out
by reorganizing and watering down documentation that is available
elsewhere. Even the content in the exemplary Part IV can be found in
several other places. It's just a lot less technical and
intimidating
here.
There are better ways to write a clusters book. I know because I've
read several cluster book outlines by members of the cluster
intelligentsia that would have been better than this offering. So
I'm
not going easy on O'Reilly, no matter how good their intentions. The
cluster community has a difficult enough time assisting people with
clusters without books like this dynamiting the proverbial cluster
well. The statement on page 28, "...benchmarking is probably a
meaningless activity and waste of time," is just plain wrong and
demonstrates a glaring lack of cluster understanding.
If you really want to learn about clusters, pick up a copy of
Sterling's "Beowulf Cluster Computing with Linux," 2nd edition, or
check out Warewulf, Rocks, OSCAR, OpenMosix, and ClusterWorld
online.
You could join a mailing list, like the Beowulf mailing list, and
subscribe to ClusterWorld Magazine. This is where the creators and
maintainers of all that is clustering hang out, announce, debate,
rant, create, lurk, help, and publish. If you want to be part of
clustering's future, then you'll check out the community's Cluster
Agenda and attend this year's ClusterWorld conference.
=================================================
Glen Otero received his Ph.D. in Microbiology and Immunology from
UCLA
in 1995 and immediately escaped to the more temperate climes and
better surf in San Diego. After some research on the molecular and
cellular biology of HIV and Herpes viruses at the Salk Institute for
Biological Sciences, Glen left the wet lab research bench in 1999.
Although leaving the research bench, he didn't leave science
altogether; traveling all the way across the street to the San Diego
Supercomputer Center (SDSC) for a stint at the Protein Data Bank. It
was while at SDSC that Glen had his Linux clusters and
bioinformatics
epiphany. Soon after that illuminating event, Glen founded Linux
Prophet, a bioinformatics consultancy specializing in the
implementation, design, and deployment of Linux Beowulf clusters in
the life sciences. Late in 2002 Linux Prophet evolved into
Callident,
a Linux cluster software and high performance computing company.
1B1B,1D1D,2626Glen Otero Ph.D.
Linux Prophet