Re: netperf & transmit path performance

From: Rask Ingemann Lambertsen (
Date: Fri Jun 02 2000 - 17:16:47 EST

Den 31-Maj-00 20:08:34 skrev Andi Kleen fĝlgende om "Re: netperf & transmit path performance":
>Kallol Biswas <> writes:
>> 57344 57344 4096 300.07 6.96 3.52 2.16 41.461

>> I am a bit concerned about the throughput over half duplex 10Mbps
>> connection. Could some body tell how throughput can go low at 10Mbps?

>Collisions of the data with the acks. The TCP ack clock gets out of sync.
>> Simon Andrew Boggi got similar result over a half duplex connection
>> with eepro100.
>> With 21143 based card we get the throughput around
>> 9.2 on 10Mbps. I do not know if there is any problem with 82559 in half
>> duplex mode.

>The 21143 uses a special patented Ethernet collision backoff algorithm that
>works better than the normal backoff. The 82559 probably doesn't.

   FWIW, on a 66 MHz 486 box, linux 2.4.0-test1, with a WD8013 board (ISA,
shared memory 8390 design), I get (netperf -t TCP_STREAM -H
-l 300 -- -s 57344 -S 57344 -m 4096 -M 4096)

Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec

114688 65534 4096 300.03 8.94

   (With the -c and -C options, I'd get the error message "recv_response:
partial response received: 0 bytes".)

   If the difference is in the collision backoff algorithm, obvioulsy more
and much older chips than the 21143 must be using one that is different
from the normal one.


| Rask Ingemann Lambertsen | E-mail: |
| Please do NOT Cc: to me or the | WWW: |
| mailing list. I am on the list.| "ThrustMe" on XPilot, ARCnet and IRC |
| Never underestimate the bandwidth of a CD-ROM flying through the lab. |

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
Please read the FAQ at

This archive was generated by hypermail 2b29 : Wed Jun 07 2000 - 21:00:16 EST