[OT] IBM in talks to buy Sun

[Arne Vajhøj]

None of which would work anything alike. And you can drop the last two
words of that sentence.

At 20 hours per months it would be 10 million man hours per month.

Provided that a few of them learn the thing well enough to provide leadershp
and oversight.

They will eventually learn. Maybe the first few versions would have
a few "untraditional features", but it is not that different from
any commercial operation where a development team is replaced (handed
over to a dedicated maintenance team or oursourced or ...). It happens.
The trip is a bit bumpy. But usually it eventually comes through. And
waste of hours does not matter for voluntary work.

Arne

 

[Arne Vajhøj]

Sounds not worth it to me.

Companies spend billions of dollars on expensive storage
systems.

I don't think they would unless they needed them.

There are now consumer RAID systems, which
often allow any (so long as all are identical) drives to be used in
them. They'll be reliable enough. Maybe you'll have to replace a drive
in the array a bit more often, but if those drives are a *lot* cheaper
the savings will add up and add up.

If it does not meet performance or maintenance criteria the low cost is
no good.

This is why IBM and Sun are in talks to possibly merge; the high-end
hardware market is shrinking, being attacked from below by commodity
microprocessors and drives and similar hardware, cheap and parallelizable.

If high end means anything above home PC grade then the high end
hardware is not shrinking.

It is changing a bit though. Solaris is getting squeezed by Linux, so
SUN has a problem and IBM does not have a problem.

(IBM has 13 billion dollars in cash to buy for)

Arne

 

[Arne Vajhøj]

It's "how badly does the owner of that 14TB disk farm want to keep it
working" that I was wondering. With hardware prices falling, the amount
you can charge them to keep it working will also eventually have to
fall. In the limit, if 14TB disk farms cost pennies and were easy to
install and use, people who needed them would just buy them, get them
running, and replace them whenever they went kaput, and the heck with
maintenance contracts. Much as they do with PCs now even though at one
time *nobody* treated a computer that way.

That is comparing apples to oranges:
1970 2010
important server must run => maintenance must run => maintenance
home PC not existing does not matter => no maintenance

Arne

 

[Arne Vajhøj]

That might change when they do the math.

They already do the math.

To compensate for slower speeds, double the number of file servers
behind a load-balancer, and ultimately the number of disks. This works
as long as you don't need to serve *single* files *really* fast.

It does not work at all.

To compensate for lower reliability (but consumer grade hardware is
getting better), assume a doubled disk replacement rate in the RAIDs.

Overall, that means four times the disks. If they're $200/TB each the
above doublings produce $800 in place of the $1000-1500 you cite for the
non-consumer-grade hardware.

Parallelism in various forms (multiprocessors, load-balanced clusters,
RAID, and so forth) make consumer grade hardware able to "add up" to be
equivalent to higher-grade hardware. Sometimes still with lower price tags.

Multiprocessors is a characteristics of non-consumer hardware.

But the rest is correct. Many cheap thingys is often less expensive
that a few expensive thingys.

And if the cheaper solution can solve the task, then it is better.

But it usually can not.

It requires specially written applications to utilize those types
of parallelisms.

Actually, I recall hearing of Sun Microsystems working on something like
that fairly recently. A plug-and-play "data center in a box" the size of
a standard shipping container. I think they planned to even rent them out.

It is called "Sun Modular Datacenter".

It exists.

But it is (usually) filled with enterprise class stuff
not consumer class stuff.

Arne

 

[Arne Vajhøj]

If anyone still isn't persuaded by Larry's argument, they should
consider Google. And i don't mean go and do a search. Google run their
data centres in exactly this way, using vast amounts of cheap commodity
hardware. It's an approach that doesn't work at the small scale, where a
single better machine plus operating costs may well work out cheaper
than two or more cheaper, less reliable ones, but at the large scale, it
works out very well indeed.

Google uses that paradigm.

But:
- Google has its own file system, its own web server and its own
back end apps to manage it
- Google usage pattern is different from most OLTP usage

The Google case proves that it can be a good solution.

It does not prove that it will always be a good solution.

I find it difficult to see a traditional eCommerce Java EE
app with a relational database at the back utilizing that model.

Arne

 

[Arne Vajhøj]

It's "how badly does the owner of that 14TB disk farm want to keep it
working" that I was wondering. With hardware prices falling, the
amount you can charge them to keep it working will also eventually
have to fall. In the limit, if 14TB disk farms cost pennies and were
easy to install and use, people who needed them would just buy them,
get them running, and replace them whenever they went kaput, and the
heck with maintenance contracts. Much as they do with PCs now even
though at one time *nobody* treated a computer that way.

The owner who wants to keep something working may consider the higher
cost a hedge[*] against loss of service or data, both of which are
typically more valuable than the hardware itself.
[*] <http://en.wikipedia.org/wiki/Hedge_(finance)>

Quite.

You can't tread a RAID as a throw-away object, to simply be replaced if it
fails, unless you regard your data with the same cavalier attitude.

The notion that you can simply hot-swap disks if they fail in a high-
volumn production environment is wacky.

It's not just data, it's service time that's valuable. Most such
installations cannot afford downtime, at least not much. Disk
failures are a major fubar. The motivation for RAID itself proves
that - it increases the cost per storage amount in favor of higher
reliability.

The suggestion that one can buy inferior hardware in quantity and just
swap in new pieces when something breaks betrays an utter lack of
understanding of the problem. If that idea worked, the people
responsible for those data centers would do that, but it doesn't and
they don't.

If you can afford to develop special software to utilize such a hardware
environment, then it can work - and it does work.

But it does not make financial sense for most companies.

Arne

 

[Mike Schilling]

At 20 hours per months it would be 10 million man hours per month.

I know. It's terrifying what sort of havoc that could produce.

 

[Lew]

Multiprocessors is a characteristics of non-consumer hardware.

While I agree with most of your points (well-supported by the facts as they
are), this is not true any more. For years more and more consumer PCs are
multi-core and/or multi-processor. You can hardly buy a new single-processor
PC any more.

<http://www.dell.com/home/desktops>
Twelve models, 48 configurations of the basic "home" models that Dell
advertises, and only one, the least expensive, has a single-core
single-processor architecture. Many have effectively eight - four cores, two
threads per core - with prices starting at a skootch over one grand.

 

[Larry K. Wollensham]

They already do the math.

Apparently not.

It does not work at all.

Sure it does.

Multiprocessors is a characteristics of non-consumer hardware.

My own PC is quad-core. Blue's is quad-core. I can't find many PCs for
sale that aren't at least dual-core, except for used and fairly old ones.

And if the cheaper solution can solve the task, then it is better.

But it usually can not.

Often it can, if done correctly.

It requires specially written applications to utilize those types
of parallelisms.

Software only has to be written once, and then it is easy to copy it and
use it wherever it's needed.

I notice an undercurrent of personal hostility in this latest series of
posts of yours. I guess my publicly calling attention to your probable
connection to Sun still bothers you? Well it's not like you made any
effort to really disguise your affiliation, though.

Of course, that affiliation may also directly explain your adamant
attitude here that consumer-grade parts will never cut it, even in the
face of the massive counterexample of Google that came to light. Sun is,
after all, in the business of selling the hardware that is beginning to
be increasingly replaced by cheaper hardware.

 

[Eric Sosman]

At 20 hours per months it would be 10 million man hours per month.

Man-hours are not additive. See ISBN 0-201-83595-9,
or if you don't want to look it up see if you can spot the
flaw in this argument: "The dentist will need half an hour
to fill that cavity for you. Since you really hate going
to the dentist and want to get the experience behind you
as quickly as possible, assign thirty dentists to the job
and it'll be done in one minute."

 

[Lew]

Man-hours are not additive. See ISBN 0-201-83595-9,
or if you don't want to look it up see if you can spot the
flaw in this argument: "The dentist will need half an hour
to fill that cavity for you. Since you really hate going
to the dentist and want to get the experience behind you
as quickly as possible, assign thirty dentists to the job
and it'll be done in one minute."

The usual cliché is, "Nine women can't have a baby in one month."

The distinction between throughput and transaction time is significant,
particularly in enterprise systems.

 

[Tom Anderson]

Apparently not.

Moreover it is *not possible* to do the maths. To do it, you'd need
accurate numbers for the failure rates of enterprise-level drives, and as
the Google paper makes clear, that data *does not exist*.

The idea that because a lot of people buy Sun/IBM/whatever hardware it
must be a good idea is pretty dodgy. The assumption is that the people who
made that decision knew what they were doing. The simple fact is that most
people, in any line of work, don't know what they're doing, and make
decisions based on a variety of factors other than what's actually best -
what they've done before, what everyone else is doing, what a salesman
told them to do, etc. That the expensive-disk option is popular doesn't
mean it's good any more than the fact that a lot of people use SQL Server
means that's good.

tom

 

[Lew]

popular doesn't mean it's good any more than the fact that a lot of
people use SQL Server means that's good.

Actually, SQL Server is pretty good.

 

[Arne Vajhøj]

Actually, SQL Server is pretty good.

Yep.

Good SQL dialect, good tools, good way of doing SP's, SP's in C# etc..

Not bad.

They even supply a JDBC driver.

That some of the design comes from Sybase does make the
design less good.

Arne

 

[Eric Sosman]

The usual cliché is, "Nine women can't have a baby in one month."

Another concerns one man digging a post hole in three
minutes, and a team of a hundred eighty doing the job in
one second flat. (If you're not sure what a "post hole" is,
it's a particular kind of one of those things you don't know
that other thing from ...)

I'd originally written "If you want an heir tomorrow,
impregnate two hundred seventy women today." But then I
concocted an alternative, hoping to forestall the objection
that the labor (sorry) of women doesn't bear (sorry) on the
topic of *man* hours. Besides, I didn't want to be a sexist.

 

[Arne Vajhøj]

While I agree with most of your points (well-supported by the facts as
they are), this is not true any more. For years more and more consumer
PCs are multi-core and/or multi-processor. You can hardly buy a new
single-processor PC any more.

You will get multi core.

Multi socket is still a characteristics of non-consumer hardware.

Arne

 

[Arne Vajhøj]

Apparently not.

If you believe that you are better doing that type of calculations
than CIO's then you should start a company in an IT intensive business
and become a multi billionaire.

When that happen then post a note here.

Until that happen then I will assume that the worlds CIO's knows
more about cost than you do.

Sure it does.

It does not.

Which is probably why that config is never used.

My own PC is quad-core. Blue's is quad-core. I can't find many PCs for
sale that aren't at least dual-core, except for used and fairly old ones.

Everybody get multi-core.

But not multi-socket.

Often it can, if done correctly.

Base on your file server example, then I do not give much for
your judgment in this area.

Software only has to be written once, and then it is easy to copy it and
use it wherever it's needed.

The majority of software is only used in one copy. So the argument does
not hold water.

Arne

 

[Arne Vajhøj]

Man-hours are not additive. See ISBN 0-201-83595-9,
or if you don't want to look it up see if you can spot the
flaw in this argument: "The dentist will need half an hour
to fill that cavity for you. Since you really hate going
to the dentist and want to get the experience behind you
as quickly as possible, assign thirty dentists to the job
and it'll be done in one minute."

True.

But since my point was that a 500000 developers team could
develop a lot of stuff in parallel, not that they could
complete a task in no time, then I don't think
it is particular applicable.

Arne

 

[Mike Schilling]

I'd originally written "If you want an heir tomorrow,
impregnate two hundred seventy women today." But then I
concocted an alternative, hoping to forestall the objection
that the labor (sorry) of women doesn't bear (sorry) on the
topic of *man* hours.

Yeah, that analogy lacks gravidas.

 

[Larry K. Wollensham]

If you believe that you are better doing that type of calculations
than CIO's then you should start a company in an IT intensive business
and become a multi billionaire.

Perhaps you should read Tom's post before you attack me again.

It does not.

It does too.

Everybody get multi-core.

But not multi-socket.

Lots of consumers have, or soon may have, multi-socket.

http://www.asc.edu/news/nov06feature.shtml mentions multi-CPU machines
being used as both servers and workstations.

Often it can, if done correctly.

[unprovoked rudeness with no useful content]

I don't like your attitude. Please either remain civil or plonk me. Your
choice.

The majority of software is only used in one copy. So the argument does
not hold water.

That's nonsense.
http://www.iht.com/articles/2009/03/17/technology/17cloud.php describes
how the kind of software used by the likes of Google to span databases
or search indices across large amounts of commodity hardware is being
open sourced and seeing increasingly widespread use. This should not be
the case unless I am right and you are wrong.

MapReduce pwn5 j00 Arne!