[135] in linux-net channel archive
TCP/IP Ethernet Benchmark
daemon@ATHENA.MIT.EDU (Leonard N. Zubkoff)
Sat Mar 18 18:19:34 1995
Date: Sat, 18 Mar 1995 11:08:43 -0800
From: "Leonard N. Zubkoff" <lnz@dandelion.com>
To: linux-net@vger.rutgers.edu
I ran some simple TCP/IP over Ethernet benchmarking last night that I thought
people might be interested in. On my home Ethernet which had no other traffic,
I used the TTCP program to test receive and transmit performance between a
486DX2-66 system with Genoa TurboExpress VL motherboard (midkemia) and my
Pentium-90 system currently using a SuperMicro P54CT-PCI motherboard.
I had originally purchased a pair of the SMC Elite Ultra Ethernet cards, but
then abandoned them last June in favor of 3COM 3C509 cards when I discovered
the problem hardware between the SMC card and bus-mastering SCSI controllers.
Since the SMC Ultra driver has been fixed to avoid this shared memory conflict
problem, I'd been meaning to try the Ultra's again. I also have one of the
BOCA Lance PCI cards, which although it sees the same "bus-master arbitration
failures" that others have reported, I still thought would be interesting to
compare with the others; no bus-master arbitration failures occurred during the
short test interval.
The test command I used was "ttcp -r -s -b 32767" on the receive side and "ttcp
-t -s -b 32767 <hostname>" on the transmit side. Here are the results of the
tests (all receive/transmit reports are from the perspective of midkemia, the 486DX2-66):
3C509 midkemia <==> 3C509 kelewan
Receive: 674.14 kb/sec
Transmit: 800.57 kb/sec
SMC Elite Ultra midkemia <==> 3C509 kelewan
Receive: 676.12 kb/sec
Transmit: 1030.52 kb/sec
SMC Elite Ultra midkemia <==> SMC Elite Ultra kelewan
Receive: 1095.38 kb/sec
Transmit: 1093.41 kb/sec
SMC Elite Ultra midkemia <==> BOCA Lance PCI kelewan
Receive: 915.42 kb/sec
Transmit: 1079.33 kb/sec
Just for the record, I also note that the SMC Elite Ultra now seems to work
quite reliably in the presence of the BusLogic BT-445S VL-BUS SCSI controller,
using the same type of test case that helped me locate the original problem.
Before the patch to the SMC Ultra driver, this combination would lead to a
corrupted file system.
Given these performance numbers, I think I'll wait until a truly working PCI
Ethernet card is available before trying again. The raw TCP/IP performance
above is awfully close to the theoretical maximum, and I'm quite impressed.
The above testing was all done on Linux 1.2.1, naturally.
Leonard
