Problems with LinkSys 10/100 cards

Robert G. Brown
Sun Feb 7 07:10:58 1999

On Fri, 5 Feb 1999, Brian Denheyer wrote:

> >>>>> "Robert" == Robert G Brown <> writes:
>     Robert> Yes, note well that measured transfer rates depend
>     Robert> strongly on CPU speed and packet size for packets less
>     Robert> than around 1K in size.  There is a graph, and set of
>     Robert> instructions for measuring performance with netperf, on
>     Robert> (the graph is even for a tulip
>     Robert> card, although not a linksys).  Measured throughput will
>     Robert> also depend strongly on the CPU load on the receiver (and
>     Robert> other things, but these three are the primary
>     Robert> determinants).
> I looked at the graph and let me tell you I am getting nothing like
> that performance.  Now supposedly the hacks which Steve posted should
> help, but since I can't get 0.90k to behave reliably, I haven't rolled
> those back in.  I am also trying to figure out what they do, and so
> far the hacks do NOT make sense in relation to the datasheet.  Pretty
> normal for hardware...  I am running these tests with very low CPU
> load, i.e., nothing else of significance is going on.

I'm curious.  I may have asked these questions before (a couple of
months ago) but for those of us with brains damaged by a hard youth the
world is always new, so bear with me.  Do you still get the same general
KIND of performance as is indicated in the graph?  That is, a region of
linear growth in speed where the NIC is sending its maximum packet rate
(so speed is varying linearly with the size of the packet payload),
followed by a saturation region where the number of packets per second
decreases as the size of the packets increases?  Where, in the case of a
normal card, the saturation occurs where the local transmission
requirements hit wirespeed and in the case of the lite-on they are being
bottlenecked by something else?

> I'm in the process now of trying to track down the differences between
> the PNIC and the 21143 and see if there is something obvious going on.
> But as of right now both netperf and netpipe break the drivers quite
> repeatably.  The problem does seem worse in "one direction".
> Of course the PNIC datasheet is *1994*, so I think I'm pretty much not
> going to be able to figure out what's going on.

I thought that the guy who posted one of the patches (Steven Huang?)
worked for Linksys.  He ought to be able to send you a current datasheet
(or provide even more substantial help).  There are a lot of linux users
out there -- my newspaper (in the article that trumpeted the fact that
Dell now officially supports linux) pointed out that linux owns 18% of
the server market and has the fastest growth rate by far of any of the
major operating systems in all markets.  Linksys appears to be wise
enough to recognize this and to maybe be making an effort to ensure that
their cards work with linux.

> *sigh*
> It'd be a heck of a lot easier to go out and buy two new network
> cards.

At $30/each, this is MY favorite solution;-).  But <sigh>, with 21140's
getting scarce, I suppose that one day having the lite-on chip work
properly would be very useful.  I can't help much with the actual
debugging, though -- I do have one remote box with a lite-on chip in it,
but it is on a 10-base hub.


Robert G. Brown	             
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525