[comp.protocols.tcp-ip] TCP benchmarking

neil@pio.gid.co.uk (Neil Todd) (04/08/91)

In the next few months I'm going to be required to benchmark a number of
IP implementations. I would like to identify some readily available
benchmarking packages, further I would like at least one of the packages
to be testing in an "industry standard" way. Any ideas ?

Neil


Neil Todd			| ..In general, it is best to assume that the
PSI%234237100122::neil		| network is filled with malevolent entities
neil@gid.co.uk			| that will send in packets designed to have
Group-ID Ltd			| the worst possible effect...

vjs@rhyolite.wpd.sgi.com (Vernon Schryver) (04/09/91)

I've encountered another "benchmark."

First try the 4.3BSD `ping -l9999999 target` and look at the packet rate.
There are commercially available UNIX workstations that will deliver the
vast majority of the zillion 98-byte ether packets separated by only 9.6
microseconds.  The point is I think there are several vendors'
workstatsions that can write at or close to ethernet media speed, <<from a
user process>>.

Second, modify the more recent 4.3BSD-tahoe or reno ping.c to flush the "."
and "\b" characters only occassionally or when things get slow, so that
`ping -f` does not spend most of its time fiddling with stdio.  Then make
`ping -f` compute a packet/sec rate to be displayed at the end.  This
produces nice numbers that seem to vary depending on the speed of the
machines on both ends.

Neither of these are very interesting benchmarks; they're just fun.  Ttcp
is better for TCP or UDP, and FTP for overall file system and network speed.


Vernon Schryver,  vjs@sgi.com

reschly@BRL.MIL ("Robert J. Reschly Jr.") (04/12/91)

>
> I've encountered another "benchmark."
>
> First try the 4.3BSD `ping -l9999999 target` and look at the packet rate.
>

   ...but before you do that, be sure your mbuf code is fixed.  On BSD
systems of early 4.3TAHOE vintage or earlier, you could get an mbuf
panic from all those unprocessed ICMP Echo Replies.  Sure surprised us.

				Later,
				    Bob

P.S.  It has been long enough to forget the exact details, but it was
caused by a mget() with DONT_WAIT flag set on a receive, or something
like that.

vjs@rhyolite.wpd.sgi.com (Vernon Schryver) (04/14/91)

In article <9104120230.aa07219@SPARK.BRL.MIL>, reschly@BRL.MIL ("Robert J. Reschly Jr.") writes:
> 
> > First try the 4.3BSD `ping -l9999999 target` and look at the packet rate.
> >
>    ...but before you do that, be sure your mbuf code is fixed.  On BSD
> systems of early 4.3TAHOE vintage or earlier, you could get an mbuf
> panic from all those unprocessed ICMP Echo Replies.


I'll take your word for that.

To really test things, you don't want the replies cluttering up the net,
causing collisions.   I forgot to mention that you want to add a suitable
entry to your ARP cache so you can blast something that won't blast back.
And the best 9.6 usec numbers need an otherwise quiet net.

Of course, these fun games would be considered denial-of-service attacks by
anyone trying to use the ethernet for real work.


Vernon Schryver,   vjs@sgi.com