Subject: War story (Unix, mainframe)
Date: Mon, 04 Sep 2000 15:20:56 -0500
From: Arnold Trembley <arnold.trembley@worldnet.att.net>
To: Cory Hamasaki <Kiyoinc@xxxx.xxx>

Cory,

(snip) 
I heard an interesting war story related to mainframe computers that I
thought you might appreciate.  I heard this story from a real old-timer,
a fellow who can remember working on IBM computers prior to the 360.

In about 1995 a company here in the metropolitan area (XXXXXXX, kind of 
a travel company) had an application that ran on an HP-9000 PA-RISC
machine.  This application was some kind of telephone billing system.
It ran once a month and processed a few hundred thousand records.  The
runtime was 28 hours.  Due to the runtime, they had to wait a couple of
days until they could close the month out, and then run the job.

Someone suggested porting the application to Unix under MVS.  The
scripts ported with very little problem, ASCII to EBCIDC.  The programs
were a bit harder.  There were ten programs, and they were modified to
use IBM packed-decimal arithmetic instead of floating point.  The data
files were the hardest to convert over.  They had to convert every
record to character format (expand the floating point numbers), convert
ASCII to EBCDIC, and then convert the numeric data to IBM packed
decimal.

The programs remained in C.  This was NOT a conversion to COBOL or any
other Language.  The mainframe version of this job ran in 18 MINUTES!
I'd calculate this as roughly a 99% reduction in run time (1680 minutes
versus 18 minutes).

I told this story to several people.  All the mainframers, including the
fellow who told me the story originally, attribute the improved
performance to mainframe I/O efficiency.  The one Unix programmer I told
this story to attributed the entire performance gain to the conversion
from floating-point to IBM packed-decimal.

I went back to my original source on this, and asked if he knew whether
or not the HP-9000 PA-RISC machine had hardware floating-point or
software emulation.  He didn't know, and hadn't considered the problem.
He agreed that floating-point performance could have affected the
outcome, and agreed it would have been more pronounced if the HP only
had software floating-point.  But he reinterated his belief that most of
the performance improvment would be in I/O.  He said he knew some Unix
internals from a few years back, and didn't think the I/O subsystem was
very sophisticated.

With kindest regards,

--
http://www.arnoldtrembley.com