Problems with LinkSys 10/100 cards

Brian Denheyer briand@deldotd.com
Thu Feb 4 14:14:18 1999


>>>>> "Steve" == Steve Huang <steve@ltnlcc.com.tw> writes:

    >> >>>>> "Steve" == Steve Huang <steve@ltnlcc.com.tw> writes:
    >> 
    Steve>     YES.  The system will never display the warning message while system
    Steve> boot up or
    Steve>     media changed.

Yes that seems to work.
 
    Steve>   Do you modified the following lines in tulip.c(v.90f)?

Yes I did.  I did the modifications in the v.90k tulip.c, but it was
very easy to check and make sure that the new code went to the right
place.  I will double-check again.

I am going to work on the "register-explanation" program today
starting with registers csr5 and csr6 since they seem to give the most
useful information.

On the 168/169 do the csr5 and csr6 registers have the same bits
meanings as the dec21041 ?

    Steve>   Could you tell me how do you measure the performance and your testing
    Steve> environment ?
    Steve> I would like to simulate it. Thanks.

Can you tell me how _you_ measure performance and I will try it on my
machines.  I think it would be very heplful.

I am running :

  Linux soggy 2.0.36 #2 Sun Jan 31 09:45:18 PST 1999 i586 unknown

I load the tulip driver as a module.  I have not tried compiling it
into the kernel.  I am NOT using any options on the cards.

I have two machines using a crossover cat5 cable (2 meters).

I ncftp from one machine to the other and transfer a 30Mbyte file.
                                                     ^^^^^^^

Potential blunder alert.  I did not take into account the disk
transfer times.  They should be small compared to the ethernet times
but not when transferring @ 1-5 Mbyte/s since that is as fast as a
disk drive.  If I take into account disk transfer times the speed goes
up to ~2.5Mbyte/s, at least this lets me know that the 100base modes
are working.

In order to make the test more reliable I have started using a smaller
test file of about 6.6 Mbytes to make disk access time smaller.  The
total transfer time is now smaller and therefore more unreliable, so I
use many trials.  Here is the performance of the disk drives :

disk 1, scsi-2  : 3.5 s for a cp (i.e. cp test test2)
disk 2, scsi-uw : 0.6s ! for a cp  

Total disk access time is 4.1s, however since we are only reading from
one disk and only writing on the other the transfer time is /2 or ~2s.

Also running the test repeatedly, which I do, takes OUT the disk
transfer times since the file will end up being cached on the machine
being read from.

The way to REALLY handle this issue is to transfer files using a
ram-disk and I am trying to figure out how to set one up.  My first
attempt failed.

After running many trials using the 6.6mbyte file I see total transfer
times of 4.5 - 6.7 s.  The transfer rate is then 1Mbyte -> 1.5Mbyte/s.
If you assume /2 transfer times for the read and write disk accesses,
which is very conservative, and then subtract 2 s from the transfer
times, then these transfer rates become : 1.4 Mbyte/s to 2.6Mbyte/s.


Brian