12 Nov
2003
12 Nov
'03
10:56 p.m.
> > > When going for Sinclair and
journalists, I always bring up the Z88
> > > at some point. Old guys may remember them (or similar machines),
> > > while younger often think it's a brand new invention,
> > Not a fool mistake. IBM recently released a server called a "z800".
> > http://www-1.ibm.com/servers/eserver/zseries/i/zseries800.jpg
> Now, that's the modern incarnation of a
REAL(tm) Computer.
> > no Z80's or Z8000's in it, I
believe.
> Nop, the zSeries is todays version of a classic
/370. The only real
> differnence to its predecessor (ES9000) is a somewhat weired 64 Bit mode
> - I have still no Idea what it is good for. Shure, the address extension
> from 24 to 31 Bit was quite helpful (8 MB user address space became
> quite tight when you already need 3 for your programm and need more
> speed due caching for 3000 concurent users), but above that, I see no
> real need. We got all bells and whistles in our data base engine, but
> still don't use more than a half a gig of mem. And even there the
> address extension of earlier architectures (swaping address spaces in
> and out of a task - aka banking:) was more than enough.
The zSeries didn't come after ES/9000. It came
after S/390 which came
after ES/9000.
You're right here.
> I think, the extension was more a marketing thing
than anything else.
Have ever tried doing data operations with data
aligned to 64-bit
boundaries on a zMachine? The difference in performance is marked.
Now, that's already true for earlyer machines, always align data to
it's size (I got no idea who ever came up with the stupid idea that
unaligned data should be allowed at all), and if possible try to align
data blocks you work in to 64 Byte or larger boundries. You'll feel
the result. Optimizing code below instruction level is kind of a
hobby to me. Especialy doing hardware related optimization without
knowing the Hardware:)
Just I still see no reason for 64Bit integers or similar (as I still
see no reason for floating point on a mainframe) the few ocassions
where 32 Bit are not enough (I found less than a dozend in 25+ years)
one can always go the classic way and use multipla 32 Bit operations.
There's only one operation (or two to be correct) where I'd whish
they had used a doubble register for result: CVB (CVD). Sometime it's
usefull to convert decimal numbers with all 15 digits into binary to
calculate evenly distributed hash codes or similar.
Maybe I'm too much of a classic mainframe guy to see the need for
such extensions. A /370 (and up) is just not ment to crunch numbers
(I'd prefer a Cray/CDC here) or do other fancy stuff, a /370 is a
work horse to shovel data around, dig thru incredible data piles
and do a hard, unglamourous every day job.
They do have instruction set additions for that.
For what? Aligning? like
AES - Align Everything Somewhere
:) SCNR
Gruss
H.
--
VCF Europa 5.0 am 01./02. Mai 2004 in Muenchen
http://www.vcfe.org/