Date: Fri, 7 Mar 2003 21:33:55 -0800 From: Wes Peters <wes@softweyr.com> To: Doug Ambrisko <ambrisko@ambrisko.com> Cc: Paulo Roberto <nirv199@yahoo.com>, Bram Van Dam <gandalfbram@pandora.be>, freebsd-hackers@FreeBSD.ORG Subject: Re: Realtek Message-ID: <200303072133.55358.wes@softweyr.com> In-Reply-To: <200303071716.h27HGHtg001578@www.ambrisko.com> References: <200303071716.h27HGHtg001578@www.ambrisko.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Friday 07 March 2003 09:16, Doug Ambrisko wrote:
> Wes Peters writes:
> | On Thursday 06 March 2003 15:02, Paulo Roberto wrote:
> | > --- Bram Van Dam <gandalfbram@pandora.be> wrote:
> | > > cheap they are they do their job fairly well. If performance
> | > > isn't an issue then go for it.
> | >
> | > I couldn't agree more. If you are just staying in 55 mph, you don't
> | > need a Ferrari.
> |
> | It's not a ford vs. ferrari problem, it's that the ford only has
> | first gear, so you're doing 45 mph at redline and in grave danger of
> | blowing the heads off continuously.
> |
> | The problem with the RealTek chipset is that the packets have to be
> | aligned on some completely stupid boundary in memory (32 bytes if
> | memory serves). The driver code ends up copying the packet data to a
> | tempory buffer before sending it for almost every outgoing packet,
> | which is just totally stupid.
>
> [snip]
>
> | JUST SAY NO.
>
> Actually, test and the pick the fastest tends to be better.
>
> Since D-Link dropped their good 4-port card for a broken one which they
> discontinued we had to scramble for a solution. Our test bed was a
> basically a "server" machine tied to a "router/bridge" like thing with
> 4 clients. We'd run tests all to the server, all to the clients and
> everything at once. This illustrated the HW issue with the new D-Link
> 4 port card since none of their "supported" drivers and OSes could get
> over 20Mbs. We had 100FDX links to each client and a Gig link to the
> server. FreeBSD could peak to 40Mbs if I recall right and we were told
> FreeBSD must be broken even though it was faster then their supported
> OSes (Windows < 1Mbs)! To be honest I did fix a bunch of bugs in the
> FreeBSD driver.
>
> Using this framework we had a bridge riser card that we could plug
> 4 various PCI ethernet cards. We tested the dc(4), fxp(4), rl(4),
> sis(4) cards of various types and with our motherboard and CPU the
> rl(4) 8139C chips where the fastest via netperf with a significant
> margin. I went into the test biased against Realtek but couldn't
> justify not using them after testing. Now we are using the 8100L chip
> so we have a pretty simple design.
You did something truly bizarre. I've tested similar cards on many
machines ranging from K6-2 400MHz to P4 2.4GHz and the RealTek
performance has always been at or near the bottom of the heap. On the
slower processors, the overhead of aligning packets shows heavily, but it
can be noticed on any system.
A number of the chips folded into the dc(4) driver are horrible and
perform right down there with the RealTek, but a few are fairly good.
The 3com 3c905s are generally quite good using the xl(4) driver, as are
the Intel EEPro's using fxp. I've read of others struggling with both
but never encountered this myself. I tend to be quite conservative about
throwing random versions of FreeBSD at systems, though, and many of those
complaints come from people at various points on -stable, rather than a
known release point.
> So I'd say given a sufficiently fast CPU and memory the Realteks work
> pretty darn good.
For a sufficient engine RPM, that escort will do 85 MPH in first gear,
too. ;^)
> The speed win could be do to a slightly better
> bus interface. That was the problem with the newer D-Link 4 port card
> in that during RX the chip would take over the PCI bus for a loooong
> time.
Yup, they're complicated beasts. For someone who is not going to work on
the drivers themselves but wants performance, I suspect buying whatever
you can get in bulk for eight dollars is not an optimal strategy.
> A sufficiently fast CPU in our case is 700Mhz Celeron which is a lot
> different then pushing 100Mbs with a P5 133Mhz.
>
> Our bigger issue is bus performance on a 32bit/33Mhz bus with 3, 4-port
> cards.
>
> To date we haven't had any trouble with them and we've shipped a bunch.
Give me 1 second and I can easily bring any of your systems to their
knees, regardless of which cards you have installed. Everything is
relative. Were you watching the system load while performing your
testing? Was the cpu doing anything but routing? Is it required to for
your application? These and many others are all important questions, and
tend to have different answers for every application. For a desktop
workstation with undemanding network application requirements (email, web
browsing, occasional software updates) RealTek or any other card that
successfully attach to the network and correctly autonegotiate with your
hub (shudder) or switch is fine. Even a RealTek. ;^)
--
Where am I, and what am I doing in this handbasket?
Wes Peters wes@softweyr.com
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200303072133.55358.wes>