Re: Tulip or De4X5 driver?

[Mike Barnes <mike barnes sci monash edu au>]

gshamov wrote:
> We have a cluster built from eight DS10L machines (under RedHat 6.2
> Linux-Alpha),linked with Fast 100mb Ethernet  switcher.

I spent a year (on and off - mostly off) getting a cluster of 7 DS10s to
work happily. Finally managed it about a month ago.

> 1) Is it possible to aquire full-duplex, 100MB network performance for
>    our hardware configuration?

I'm guessing there's not much difference between the DS10 and DS10L, so
I'd say "yes".

> 2) Which network card driver gives better performance -- Generic
>    DE4X5 or Tulip? And how to configure these drivers for maximum
>    network speed? We have experimented with both of them, using
>    kernels 2.2.18 and 2.4.2 -- it seems that its impossible to have
>    Full-Duplex with DE4X5... But it seems that network with FDX Tulip
>    is slower than in case of Half-Duplex De4X5... What shoulde we do
>    wit it?!!

We settled on the Tulip, eventually.

> May be someone there had such a experience with these DS10L systems?
> Any suggestions will be gratefully appericated!

I'm going to start by guessing that you have a switch that tries to
autosense link speed, am I right? If your switch ports can be manaually
forced to run at 100Mbps/Full Duplex, then do so and it might all be
fine right now. Otherwise, here's the distilled wisdom ...


1) Install the latest firmware update.

2) In SRM, set ewa0_mode (or whatever adapter) to "FastFD".

3) Fix the Tulip driver! Stephen Degler sent the following to this list
on January 29th this year. It was not applied to the 2.4.1 kernel
sources, I don't know about 2.4.2. I _think_ the same problem exists in
2.2.x kernels in a similar form, but I haven't looked lately.

> This one-liner fixes a subtle 21143 autonegotiation problem for me on a Zynx
> quad card.  The driver would claim to negotiate 100-FD, but would report late
> collisions and bad transmit throughput.
> 
> The driver still allows packets to be transmitted during autonegotiation,
> but that only drops a few packets.
> 
> skd
> 
> --- 21142.c.bad Sun Jan 28 15:26:25 2001
> +++ 21142.c     Sun Jan 28 11:51:59 2001
> @@ -171,7 +171,7 @@
>                         for (i = 0; i < tp->mtable->leafcount; i++)
>                                 if (tp->mtable->mleaf[i].media == dev->if_port) {
>                                         tp->cur_index = i;
> -                                       tulip_select_media(dev, 0);
> +                                       tulip_select_media(dev, 1);
>                                         setup_done = 1;
>                                         break;
>                                 }

Give that lot a try. Feel free to email me direct on this - I know what
you're going through. :)

Mike.