On Sun, Nov 28, 2010 at 1:56 AM, David Masover <ninja / slaphack.com> wrote:
>
> In other words, you need someone who will support it, and maybe someone who'll
> accept that kind of risk. None of the Linux vendors are solid enough? Or is it
> that they don't support mainframes?

Both, and the Linux variant you use has to be certified by the
hardware vendor, too. Essentially, a throwback to the UNIX
workstations of yore: if you run something uncertified, you don't get
the support you paid for in the first place.

> "Both the original Tianhe-1 and Tianhe-1A use a Linux-based operating
> system... Each blade is composed of two compute nodes, with each compute node
> containing two Xeon X5670 6-core processors and one Nvidia M2050 GPU
> processor."
>
> I'm not really seeing a difference in terms of hardware.

We are probably talking on cross purposes here:
You *can* build a vector CPU cluster out of commodity hardware, but it
involves a) a lot of hardware and b) a lot of customization work to
get them to play well with each other (like concurrency, and avoiding
bottlenecks that leads to a hold up in several nodes of you cluster).

> Sorry, "unproven, unused, upstart"? Which language are you talking about?

Anything that isn't C, ADA or COBOL. Or even older. This is a very,
very conservative mindset, where not even Java has a chance.

> So it very likely does pay off in the long run -- being flexible makes good
> business sense, and sooner or later, you're going to have to push another of
> those 15-month changes. But it doesn't pay off in the short run, and it's hard
> to predict how long it will be until it does pay off. The best you can do is
> say that it's very likely to pay off someday, but modern CEOs get rewarded in
> the short term, then take their pensions and let the next guy clean up the
> mess, so there isn't nearly enough incentive for long-term thinking.

Don't forget the engineering challenge. Doing the Great Rewrite for
software that's 20 years in use (or even longer), isn't something that
is done on a whim, or because this new-fangled "agile movement" is
something the programmers like.

Unless there is a very solid business case (something on the level of
"if we don't do this, we will go bankrupt in 10 days" or similarly
drastic), there is no incentive to fix what ain't broke (for certain
values of "ain't broke", anyway).

> Also, think about the argument you're using here. Why hasn't it been done? I
> can think of a few reasons, some saner than others, but sometimes the answer
> to "Why hasn't it been done?" is "Everybody was wrong." Example: "If it was
> possible to give people gigabytes of email storage for free, why hasn't it
> been done?" Then Gmail did, and the question became "Clearly it's possible to
> give people gigabytes of email storage for free. Why isn't Hotmail doing it?"

Google has a big incentive, and a big benefit going for it:
a) Google wants your data, so they can sell you more and better ads.
b) The per MB cost of hard drives came down *significantly* in the
last 10 years. For my external 1TB HD I paid about 50 bucks, and for
my internal 500GB 2.5" HD I paid about 50 bucks. For that kind of
money, you couldn't buy a 500 GB HD 5 years ago.

Without cheap storage, free email accounts with Gigabytes of storage
are pretty much impossible.

CUDA and GPGPUs have become available only in the last few years, and
only because GPUs have become insanely powerful and insanely cheap at
the same time.

If you were building the architecture that requires mainframes today,
I doubt anyone would buy a Cray without some very serious
considerations (power consumption, ease of maintenance, etc) in favor
of the Cray.

-- 
Phillip Gawlowski

Though the folk I have met,
(Ah, how soon!) they forget
When I've moved on to some other place,
There may be one or two,
When I've played and passed through,
Who'll remember my song or my face.