Subject: Re: ballpark ratio users:cpu for a 86 NetBSD box?
To: None <firstname.lastname@example.org>
From: John F. Woods <email@example.com>
Date: 09/28/1994 09:53:48
This is getting way off topic, but I'd like to make a couple of observations,
if for no other reason than to cheer up the owners of non-Intel systems :-).
Running a big multi-user system isn't just a matter of CPU speed, unless
all your users are doing nothing but using xv to look at JPEG images.
Delivered disk bandwidth is a crucial part of *most* multi-user mixes,
and that's something that expensive workstations have generally supplied
much better than PCs have (and PCI or no PCI, I'll believe that a given
PC has a *controller* that delivers better performance when I see a real
benchmark run on it).
As an unusual example of this effect, when I brought up my NetBSD system
full time, it was replacing an ancient CRDS Universe system, which contained
a 12.5MHz 68020 central processor (and a 10MHz 68000 handling 12 serial
ports). According to dhrystone, the new system was approximately ten times
the horsepower of the old system (and the memory system was at least four
times the speed, too, another important thing that often gets lost in
clock-size wars). Yet as purchased, my 486/33 had great difficulty keeping
up with the Universe system doing 38.4KB SLIP, and the Universe was still
handling the modem at the time. (There have been improvements in NetBSD's
driver software since then, so *maybe* it would now be a fair match :-).
And, of course, if I had actually been using all twelve serial ports, it
would have been quite difficult to have matched the capability of the
Universe. (And if I had maxed out the Universe at 68 ports?) Horsepower
just isn't enough to make up for cripplingly braindamaged I/O design.
Fortunately for the PC side of this argument, the CRDS Versabus SCSI channel
board was really slow, so my disk bandwidth improved quite a bit :-). Though
not by a factor of ten.