> NE2100 clone on VLB using le0, 486/100 -> DEC 5000/240 Ultrix
> over busy mildly ethernet (10-15% base load) 400kbytes/s

This seems way too low.  Ultrix machine must be the bottleneck.

> One thing I always do is "put junk /dev/null" several times to get
> an idea of the maximum throughput.  Some cards seem to have problems
> when you are writing to disk while the network is active. (which is

use ttcp to get an idea of maximum throughput. ttcp will
avoid file system overhead + you can control a number of
parameters.  You can get ttcp from ftp.sgi.com:sgi/src/ttcp.
Sample use:

ttcp -r                         # on the receiver machine
ttcp -t <receiver-machine-name> # on the sender machine

With ttcp I get about 2MB/s when the sender + receiver are
on the same machine (a 25Mhz 486) & 285KB/s from this PC
(with WE8013EPC) to a Sun3 (sunos3.5).

What kind of loopback throughput do people see on a PCI
machine?  By loopback I mean IP loopback (sender/receiver on
the same machine).  For comparison, on a SGI R4000/100Mhz
Indigo2 I have seen loopback numbers in the neighborhood of
26MB/s.  Even an R4000/100Mhz Indy is not too shabby
(atleast 11MB/s).

Bakul

PS: of course, anything about 2 MB/s in loopback mode is
wasted if all you have is a 10Mbps interface like ethernet
but the loopback numbers do tell you something about the
system overhead.