[Info-vax] The Future of Server Hardware?

[Johnny Billquist]

On 2012-10-02 19:20, Stephen Hoffman wrote:
> On 2012-10-02 16:32:12 +0000, JF Mezei said:
>
>> Since I am good at asking stupid questions...
>>
>>
>> Assuming, for the sake of discussion that todays's x86 , Power and the
>> architecture that started as the 360 have equal CPU horsepower.
>
> Mainframes typically excel at providing high bandwidth and low latency
> into and out of the processors, and not strictly at raw CPU speed.
> Though they're typically no slouches there, either.
>
> Mainframes are centrally designed for error detection and correction;
> what HP calls RAS.
>
> And with the IBM boxes I've worked with, the I/O channels are gonzo
> fast, too.
>
> The older IBM mainframe boxes I've used were painful to set up - you had
> to specify everything, and in detail - but once you had the jobs
> configured, even the small mainframes were silly-fast.  Batch was
> screaming fast.  (Why "older"?  I haven't had to configure DASD and the
> other related tasks in many years, and I suspect that IBM has
> substantially improved here.)
>
> Few folks will choose to run large-scale web servers and web sites on
> mainframes for that matter, though it is technically possible.  Other
> than those folks with very specific requirments or that have a
> willingness to spend far more money on iron than the average, that is.
> There are other (and usually more cost-effective) ways to serve (for
> instance) web requests, though you might well find a mainframe as a
> back-end somewhere.
>
> Mainframes are good for some jobs.  Roadrunner or Jaguar at other jobs.
> Racks of x86 or ARM boxes at others.  And yes, Itanium has its place
> here, too.

An additional comment on what mainframes (traditionally) were good at 
was offloading. The CPU does very little on a mainframe, apart from 
actual computing. You would have intelligent I/O processors that do I/O 
for you, intelligent terminal controllers that handles everything 
related to terminals, and so on... Most modern, small computers let the 
CPU do all the work (something DEC championed).
You no longer have dedicated I/O processors, terminal processors, and so 
on...

On the other hand, this whole design philosophy have come back somewhat 
in the form of more and more offloading of work on ethernet controllers, 
graphic controllers, and so on.

Mainframes, back in the old IBM days (or DEC-10 for that matter) handled 
many users party because the CPU didn't have to be bothered just because 
a user was typing on his terminal.

But nowadays, mainframes main reason for existing is legacy. There is so 
much software around that you still want to run, and which you want to 
run on newer, faster hardware. However, developing something new from 
scratch on a mainframe? I doubt anyone would. It would be way cheaper to 
buy a PC or three, with faster CPUs and loads of cheap disk. It might 
not be able to connect to the same amount of disk at high speed, but 
really few actually need *that* much disk on a single CPU.

	Johnny

-- 
Johnny Billquist                  || "I'm on a bus
                                   ||  on a psychedelic trip
email: bqt at softjar.se             ||  Reading murder books
pdp is alive!                     ||  tryin' to stay hip" - B. Idol