Subject: RE: Why is Samba so much slower on NetBSD than FreeBSD?
To: None <bouyer@antioche.lip6.fr, CVETTE@borders.com, rhialto@polder.ubc.kun.nl>
From: Olaf Seibert <rhialto@polder.ubc.kun.nl>
List: current-users
Date: 10/23/1998 11:43:53
"Calvin Vette (IT- Borders Online)" <CVETTE@borders.com> wrote:
> Once I decoded your names :-) - 

;-) The theme here is (mainly dutch) polders. That's what you get when a
company is called Polderland.

> You've got NetBSD running on the slowest machines, with ISA Ethernet cards.
> This is going to make a huge difference. And while you didn't mention it,
> the amount of RAM could make a difference in low memory conditions. My point
> is, it's not a very scientific comparison.

I know, but it's the machines I have to test with. And I thought that
raw machine speed would not be too much of influence anyway since,
according to top, samba uses only a very small part of available CPU
anyway. Perhaps it depends a lot on peak performance though, I can't see
that with top of course. Same with raw TCP speed: that seems ok when
tested separately.

> I'd recommend testing a little more scientifically to get a real comparison
> - try swapping the NetBSD 1.3.2 drives into the NT box (Proc & RAM). 

Unfortunately the NT box is a "production" machine so I can't touch it.
(And I'm too afraid that NT will break in some mysterious way so I'm not
even going to try.).

> If it's just raw networking speed you're after, try using NetBSD current
> with UVM. (A release since the beginning of September). Chuck's pretty much
> eliminated excessive internal copying, which goes a long way towards
> improving Network I/O (and in UVM's case, all I/O).

I can try to put this on the purmer machine (the fastest one); it has
disk space left over. I can also try (as suggested below by Manuel
Bouyer's comments) to put its ne2000 clone card into polder (the other
and most important production machine) some time after hours.. Together
this should provide interesting additional info, I think.

Manuel Bouyer[SMTP:bouyer@antioche.lip6.fr] wrote:

> I would tend to blame the 3c509 for this. I've got tons of problem with
> them,
> because of their really small fifo. Could you try another ethernet
> board (ne2000 are good, but a DMA-capable board would be better for
> a server).
> In a general manner, I now tries to avoid 3com networks adapters (althouh
> their other networks products seems good).

I always thought the 3c509 were pretty good cards. So one learns something
new every day. The raw TCP speed, even from the slowest of these machines,
tends to be pretty good though. 

-Olaf.
--
___ Olaf 'Rhialto' Seibert - rhialto@polder.ubc. ---- Unauthorized duplication,
\X/ .kun.nl ---- while sometimes necessary, is never as good as the real thing.