Mainframes vs clusters

[Brian Densmore]

> From: Mike Coleman [mailto:mkc at mathdogs.com]
> pretty quick box.  One day, just for fun, I benchmarked the 
> two systems just
> to see how much slower my anemic laptop was.  I was quite 
> shocked to discover
> that the laptop was the faster of the two, by a significant 
> margin (maybe 2x,
> I forget).
> 
> The moral for me was to be wary of my guesses about the 
> relative speeds of
> hardware (in light of Moore's law).

Heh, heh, heh! As I said in an earlier post. I am currently heading up a
mainframe migration project.
We are porting a quad processor "state-of-the-art" Unisys mainframe to a
dual processor Wintel.
On average we are seeing 3x speed increase on the Wintel.
One process that used to take 3-5 hours, runs in less than 10 minutes.
One nightly stream (when the load is very low) that ran for 4 hrs runs in
less than 1/2.
Another process that runs on a weekend (because it takes 18 hrs or more)
runs in a few hours now.
But, on the flipside, there are some things that are transparent to the
mainframe which are extremely difficult to duplicate on PC/Windows
hardware/software. Although it would be simple to emulate on PC/Linux. Linux
unfortunately isn't an option at this point because there is no Unisys
mainframe emulator for Linux.
There are some things mainframes are absolutely needed for ... today. But as
Linux makes its way into the mainstream, that will be less and less true.
Technology is at a point where raw speed isn't the issue. It's how you route
that speed, how long your data and transmission paths are, and how you
utilize that speed. That's why the fastest Pentium IV is still outperformed
by much slower AMD K7s. Mainframes have wider I/O paths, but they are 1000
longer or more. One thing that will turn the scales back in favor of
mainframes is optical data paths. Mainframes are also capable of supporting
a large number of users and processes, but almost every office these days
have PCs and that power really isn't utilized or needed.