[tulip] Terrible throughput of KTI 221TX NIC (Intel 21143 chip) in fdx mode

Mon, 26 Feb 2001 09:53:56 +0100

Dear tulip experts,

  I have a beouwulf type cluster of workstations.  The nodes
have 3com 905c and KTI 221TX/3 NICs which are connected with
a KTI switch.

  The performance of the cluster was terrible, so I made some
tests with netpipe.

  The 3com cards works perfect in fdx mode.  Transfer rate is
saturating at 90Mb/s as function of the blocksize.

  However, the KTI cards show a sudden drop of the bandwith
from 60Mb/s to 1-2Mb/s at around the blocksize of 5k in full
duplex mode.  Half duplex is OK.

  Do you have any ideas what can be the reason and how to
overcome this?

  More details: I used the tulip.o driver supplied with RH7.0,
more correctly, with RH7.0's kernel upgrade to 2.2.18. I also
tried the de4x5.o driver, with similar results.  Is there any
significant difference between these two drivers?

  And a last question:  What is flow control?  My switch has
this parameter for each port, which seems to be enabled if the
port is in autonegotiation mode and disabled if in forced modes.

  Thanks for any help:

					Tamas