NIC Throughput Results

   The following pages show the results of some throughput tests I ran on a handful of NICS that I had. Don't read too much into the results as they may be specific to my setup, but they should be of interest.
   I'd been curious for a while about the effectiveness of the NICs I was using on the LRP router that I connect to Optus@home with and it turns out that I was using the worst performer (SuperLan-2U) and a close running third worst (addtron).
  Mind you they are still well above the current Optus Internet speeds!

   Manfred's Nic Performance Measurements page was the starting point for this project. I couldn't however get the tcpserver/tcpclient to give similar results. It consistently had approx 20 seconds of overhead before it would send the data so it didn't accurately show just how much could be shovelled through the link per second.
   The kernel documents had reference to the ttcp program for throughput tests, its history is here , so ignoring compilation errors I used that to give what follows.

--= HOW =--

I did all the tests via scripts from one machine (the receiver) over an rsh link. The script then parsed and sorted the data into an acceptable form for gnuplot to deal with. This is what it's based from .....

On the transmitter use:-

ttcp -t -s -p4321 -l8192 -n1024 remote_receiver

where remote_receiver is the name of the other machine
After that run the following on the remote machine (this order of execution does matter)

ttcp -r -s -p4321

A few seconds later, some blinking lights and the results will then be displayed. The following are those for the transmitter, a corresponding set will appear on the receiver.

ttcp-t: nbuf=1024, buflen=8192, port=4321
ttcp-t: socket
ttcp-t: connect
ttcp-t: 0.0user 0.3sys 0:07real 4% 0i+0d 0maxrss 0+1pf 0+0csw
ttcp-t: 8388608 bytes in 0.330000 CPU seconds = 24824.242424 KB/cpu sec
ttcp-t: 8388608 bytes in 7.403582 real seconds = 1106.491425 KB/sec

--= INTERPRETATION =-

My description of the cards either relates to the Manufacturers Name as extracted from the arpwatch programs database, or the kernel driver (depca, ewrk3, lance).
With the "combo" results the 486 cards are the shorter description eg:-"depca" The p120 are the longer description ie:- "depca de202"
user, sys, real and % have been plotted along with CPU secs and throughput and lastly - the whole reason for the exercise - Duration with Measured Throughput.
The first graph and table on each page are those of the Receiver which are the results I'm principally interested in.
As with Manfreds page I've converted the KB/sec to Mbit/sec in the following tables. This is why the discrepancy appears between the table and chart values for throughput, the cahrt uses base16 (1024) and the table base10 (1000) for kilo..

The dx66 results are from a second 486 board which is unusual in that it takes 16 sims. It currently carries 16x1 megs sims and acts as the LRP router. The bus clock speeds are also adjustable and that is what the "clk2-2" refer too. I was curious if it would make any difference.
"clk2-2" is the bus clk/2 and io wait x2
"clk4-4" is the bus clk/4 and io wait x4

The RECIEVER and THROUGHPUT results are those of most interest - I'm more interested in what can be pulled from the pipe than pushed in it.

--= CONCLUSION =--

Make of it what you will -- but....
The lance card was a stand out performer, followed by the accton.
The lance card uses the amd79c960kc chip and employs Bus Mastering, this link will explain it better than I can.  The "lance" card is actually an Eagle Systems International NFX4-T supplied by an Australian company in Moorabbin, Victoria which I assume is now defunct - but I could be wrong!
   It now sits on the firewall internal interface and should be a nanosecond or two quicker for my DNS lookups. *cough*

The comparison between the p120 and 486 is also interesting.   The increase is either the increased processor speed, available memory or board architecture. It's more than likely just the processor speed though, judging by the CPU usage.
The depca is a DE202 and has been a reliable card for me, but it ranks last in the throughput results, hogs the CPU and also has an annoying 30 second self check routine at boot up. I can use the lance with a clear conscience

The comparison between clock speeds on the dx66 seems a case of "win some, lose some". Not much change for the lance (understandable with Bus mastering); the accton responded to the change noticably but the others were minor.  The effect of bus speed changes seems to be variable and card specific.
I can't remember why the depca and ewrk3 only have one result each, either it hung the bus or I got tired of rebooting :-) Ooops.
As for the 486-66 100 comparison, nothing outstanding on the graphs; the dx4 100 wins the race with the accton having a bad hair day!

So .. if your still here and can stand a few decimal points then set your filters in place and follow the links.
The throughput results (last two links in each section) are the most relevant