RE: Think & Straight Steer & Liners

["Farmer" <pharmer@domain.elided>]

> I wouldn't be so sure about marketing hype.... there's no linux
> PR department.

You are right.  BUT - there are so many folks who want to see it succeed
that they deliberately push this kind of info.  Has Computerworld published
a report in which a Sequent Numa is compared to an SP2 lately?  Not that I
remember - but it's a much more important comparison!

> This is only a customer testimonial.  Also, I don't think a multi-billion
> dollar oil company would invest a dime in a system that they didn't think
> would be reliable, no matter what the potential reward.

Actually, in my experience the folks that usually publish this kind of thing
are senior managers that want some publicity.  Either to help get
recognition within the company, or outside of the company.  I run into that
all the time.  I worked directly with the folks on the biggest warehouse at
MCI - about 5 tbytes on a 120 node SP2.  They spent half their time chasing
down publicity.  Did they build a good product?  Well, they convinced a lot
of folks at the time, but now everyone is using it as a textbook example of
what not to do (chase technology & ignore the business problem).

> Consider also that linux supports Intel SMP systems that can be clustered
> in MPP arrays for even more power.  I've read about just that and they're
> 'out-thinking' Cray systems.

We were asking IBM for that capability about three years ago.  The salesfolk
didn't understand - but the backroom technitions were working on it.  It's a
lot of work to try to balance a database in that environment - but it has to
be done.  It's also essentially what NUMA is.  I'd be skeptical about
believing that they are out-thinking Cray.  I just don't believe that
there's a sufficiently organized amount of intellectual capital to pull that
off.

> Yes, it's computational, but so what?  Databases are being developed
> by Oracle, Informix, etc.

But running a very large database (>500gb) the SMP model runs out of steam.
Primarily due to backplane problems (the lost of time involving in having
all the cpus need to cross a longer pipe to get to whatever).  So, most big
databases (as of a few years back) go onto MPP, or nowadays, are going onto
NUMA.  Clustered SMPs only work for certain applications - where your data
can be partitioned by cluster node; otherwise you'll loose performance
bigtime when crossing the cluster network link.  To make a MPP or NUMA
platform work well, you *must* have a powerful toolset.  These toolsets are
where much of the actual innovation is.  They handle the load balancing,
monitoring, etc.  As an example: all nodes on the SP are connected to a
control workstation via RS232 and an internal ethernet.  From the control
workstation (or remote x-window display) you can see the exact status of all
nodes (remember - you could have as many as 250+ nodes).  It will help you
discover unusual performance, heat problems, warn of disk exceptions, etc.
Tools like this make or break the application.  Not to mention
backup/recovery tools and high availability.  The big databases work very
closely with the hardware manufacturer to ensure compatibility in these
environments, and for these tools.  Maybe linex will be up there in four
years - but maybe not.  The cutting edge stuff will probably be going to
64-bit unixes.

> I've been running Linux systems for almost two
> years with *no* downtime, except for that caused by my own maintenance
> schedules.  The platform is obviously just about there and that
> is evident by
> the sheer number of applications that are being written for or
> ported to the
> Linux platform.

I'm not surprised.  But - when you put an operating system into a situation
in whic h you are clustering a large number of nodes it's an entirely
different problem.  I'm looking forward to running Informix here at the
house on a small linex box - and don't really expect many problems.  *other*
than having to get rid of a ton of hardware that isn't supported (like my
cpu!)

Eric wrote:
>98% - geez I think even NT can do better than that. I don't know the
>reliability numbers for the cluster systems (Beowulf being one). I hear
>running SMP has gotten very reliable though. I don't know of putting big
>databases on a Linux cluster (I spoke with a guy in Italy who was doing
>that but don't know how it went) I've never used a cluster or even SMP,
>but on a PowerPC 604E running AIX 4.1 and thrn 4.2 I had nowhere near 4
>nines reliabity. That is one reason (including support and availability
>of drivers) that I switched to Linux on my latest project. Hey, the cost
>was nice too, but I remember how long it took me to compile a list of
>competent engineers at various IBM and Motorola sites to ask questions
>about the deep dark recesses of AIX (but I do miss CDE). Now with Linux
>I can post a question and probably get an answer in a couple of days
>from someone who wrote the driver or the part of the kernel I care about
>(and if necessary I can go poking around in the kernel source).

After running web servers on both NT and Sun, my experience was that Sun was
cheaper by far: in terms of labor.  I ran a stack of tandem NT smps.  Good &
expensive stuff.  But I don't think NT is up to mission critical
applications yet.  Although the security is better than Unix, it's way too
easy for one application to hose another.  I think it interited too much
from the pc mentality (ex: why would you want batch configuration
files?!? ).  For about the same money I think a sun server could probably
outperform an nt server 4:1 (no data to back me up, just antecdotal
comparisons).

But when it comes to the reliability of popular alternatives, I'd place MVS
first, and AIX second.  To address your comment above, Eric, if you are
running AIX on a small box (a C10, etc), you will not get the same response
that we got running a 2 million dollar SP.  IBM does most of their support
over the telephone, and beats almost everyone at it.  When I had a serious
problem with a magstar drive, I got one of the engineers who designed it on
the telephone.  Further, if I have a serious problem, I'd have it fixed
within 24 hours.  Even if IBM had to bring in a dozen folks & a big old box
of hardware.  This service is expensive, but competively priced for
important applications.

Personally, I'd love to see Linux surge ahead.  BUT - I work with a lot of
projects that fail due to oversold technology.  And personally, I feel that
the claims being made these days are being made by folks who are overeager
for success.  Reminds me a little of when Windows 3.0 hit the market and
everybody gushed about how "intuitive" it was.  What a load of crap.  As
long as you get the straight dope you can often make a solution with flaws
work just fine.  But I don't want to be surprised with that kind of a
problem 4 weeks prior to a big deadline!

Ken Farmer