[Gllug] Observed speed of gigabit ethernet only about 1/4 or 1/5 of advertised

[Peter Joanes]

Hello,

On Saturday 14 July 2007 14:52, Richard Jones wrote:
> Surprisingly with ttcp it's slower -- 170 Mbps, or about one fifth of
> the theoretical max.  With UDP it's little better -- 195 Mbps.

With the following set-up (machine B is a thin-client which is connected via a 
crossover cable),
Machine A:
	eth1: Broadcom Corporation NetXtreme BCM5704 Gigabit Ethernet
	2x Opteron 242, 2GB RAM

Machine B:
	eth0: Broadcom Corporation NetXtreme BCM5701 Gigabit Ethernet
	 VIA Epia with 800Mhz C3, 256MB RAM

I get:
# ttcp -t -s -p 631 -n 100000 -l 65536 -f m 192.168.2.7
ttcp-t: buflen=65536, nbuf=100000, align=16384/0, port=631  tcp ->192.168.2.7
ttcp-t: socket
ttcp-t: connect
ttcp-t: 6553600000 bytes in 121.52 real seconds = 411.47 Mbit/sec +++
ttcp-t: 100000 I/O calls, msec/call = 1.24, calls/sec = 822.93
ttcp-t: 0.0user 3.8sys 2:01real 3% 0i+0d 0maxrss 0+17pf 121625+56csw

# ttcp -t -s -p 631 -n 100000 -u -f m 192.168.2.7
ttcp-t: buflen=8192, nbuf=100000, align=16384/0, port=631  udp ->192.168.2.7
ttcp-t: socket
ttcp-t: 819200000 bytes in 12.80 real seconds = 488.24 Mbit/sec +++
ttcp-t: 100006 I/O calls, msec/call = 0.13, calls/sec = 7812.24
ttcp-t: 0.0user 1.2sys 0:12real 9% 0i+0d 0maxrss 0+3pf 8822+15csw

These figures would be far lower if I weren't using jumbo frames because the 
processor in machine B wouldn't be able to handle the rate of packets (CPU 
usage is ~55% during the test above).

-		Pete.

-- 
Gllug mailing list  -  Gllug at gllug.org.uk
http://lists.gllug.org.uk/mailman/listinfo/gllug