From owner-freebsd-hackers  Mon Apr  3 01:28:44 1995
Return-Path: hackers-owner
Received: (from majordom@localhost) by freefall.cdrom.com (8.6.10/8.6.6) id BAA00256 for hackers-outgoing; Mon, 3 Apr 1995 01:28:44 -0700
Received: from godzilla.zeta.org.au (godzilla.zeta.org.au [203.2.228.34]) by freefall.cdrom.com (8.6.10/8.6.6) with ESMTP id BAA00247 for <hackers@FreeBSD.org>; Mon, 3 Apr 1995 01:28:28 -0700
Received: (from bde@localhost) by godzilla.zeta.org.au (8.6.9/8.6.9) id SAA14278; Mon, 3 Apr 1995 18:22:29 +1000
Date: Mon, 3 Apr 1995 18:22:29 +1000
From: Bruce Evans <bde@zeta.org.au>
Message-Id: <199504030822.SAA14278@godzilla.zeta.org.au>
To: davidg@Root.COM, luigi@labinfo.iet.unipi.it
Subject: Re: NE2000 Plus performance
Cc: hackers@FreeBSD.org
Sender: hackers-owner@FreeBSD.org
Precedence: bulk

>1) how much faster are memory-mapped cards wrt io-mapped ones ? Is it
>   just the clock cycle per word that you save in transferring data
>   with MOVSW instead of INSW, or there is more (e.g. the driver

That is about the only difference for mine.  Much more than one clock
cycle (at 8MHz) is saved.  The memory mapped interface runs at about
3MB/sec while the i/o mapped interface runs at about 2MB/sec.  Right
now for an Addtron(?) AE-200JL in NE2000 mode on a 486DX2/66-VLB while
running ttcp -r, systat reports the following overheads:

	17.5%Sys  81.9%Intr  0.5%User  0.0%Nice  0.1%Idl

i.e., the system is saturated.  For the same setup except in WD8013EBT
mode on a 486DX/33-ISA, the overheads are only

	20.7%Sys  52.9%Intr  1.2%User  0.0%Nice 25.2%Idl

This shows that NE2000 mode has more than 29% overhead for copying
data from the ISA bus alone (the ISA bus is approximately equally
slow for equivalent i/o's on the two machines).  The interrupt overhead
won't be much lower than 52.9% on faster (ISA) machines.  I think wait
states limit the MOVSW speed to 3.3MB/sec on my systems, so 1100K/sec
for ethernet input must cost 33%.

Bruce