network performance

I did some performance testing on a network consisting of 12 machines and a
highperformance cisco switch. All machine are uring Intel network care
(82557 driver). The bulk of the traffic is comming from 6 machines all
sending that to one machine.

To simulate worst case scenario I had the 6 machines send data as fast as
possible to the server. To my amazement the server was handling a little
over 10MBYTES per secondes, hence very very close to theoretical limit. At
first I didn’t beleive it :wink: So I used somebody else test program. just to
make sure there wasn’t anything wrong with the way I did the test which was
basicaly to open ("/dev/null" ) and write() to it. The other test (based
on Send/Receive/Reply) gave the same results.

It’s the first time I got the see 100Mbits ran that fast. The switch is
doing a great job but the OS, Net and Net driver must be doing something
right…

Just though it was worth a post :wink:

  • Mario

And you’re surprised Mario? QNX4 is that fast. I’ve always been able
to run 10-base-T right up against the limit. Back when I was doing
this kind of performance testing, the CPU’s weren’t yet where you could
get 100 Mbps. But now, no sweat.

Have you duplicated this on QNX6?

Short tangent:

The enemy of ethernet performance is collisions. To get the most out
of 10 Mbps ethernet, we had dual network cards in a file server system.
All of the workstations used one network for traffic inbound to the
file servers and the file server used the other network for traffic
outbound from the file server. The file server never (well rarely) saw
a collision. And there wasn’t that much traffic from the workstations
to the server.

Mario Charest postmaster@127.0.0.1 wrote:

MC > I did some performance testing on a network consisting of 12 machines and a
MC > highperformance cisco switch. All machine are uring Intel network care
MC > (82557 driver). The bulk of the traffic is comming from 6 machines all
MC > sending that to one machine.

MC > To simulate worst case scenario I had the 6 machines send data as fast as
MC > possible to the server. To my amazement the server was handling a little
MC > over 10MBYTES per secondes, hence very very close to theoretical limit. At
MC > first I didn’t beleive it :wink: So I used somebody else test program. just to
MC > make sure there wasn’t anything wrong with the way I did the test which was
MC > basicaly to open ("/dev/null" ) and write() to it. The other test (based
MC > on Send/Receive/Reply) gave the same results.

MC > It’s the first time I got the see 100Mbits ran that fast. The switch is
MC > doing a great job but the OS, Net and Net driver must be doing something
MC > right…

MC > Just though it was worth a post :wink:

MC > - Mario

“Bill Caroselli” <qtps@earthlink.net> wrote in message
news:bqdk9f$r57$1@inn.qnx.com

And you’re surprised Mario?

Yes I am Bill.

QNX4 is that fast. I’ve always been able
to run 10-base-T right up against the limit. Back when I was doing
this kind of performance testing, the CPU’s weren’t yet where you could
get 100 Mbps. But now, no sweat.

Have you duplicated this on QNX6?

Haven’t tried.

Short tangent:

The enemy of ethernet performance is collisions. To get the most out
of 10 Mbps ethernet, we had dual network cards in a file server system.
All of the workstations used one network for traffic inbound to the
file servers and the file server used the other network for traffic
outbound from the file server. The file server never (well rarely) saw
a collision. And there wasn’t that much traffic from the workstations
to the server.

Mario Charest postmaster@127.0.0.1 wrote:

MC > I did some performance testing on a network consisting of 12 machines
and a
MC > highperformance cisco switch. All machine are uring Intel network
care
MC > (82557 driver). The bulk of the traffic is comming from 6 machines
all
MC > sending that to one machine.

MC > To simulate worst case scenario I had the 6 machines send data as
fast as
MC > possible to the server. To my amazement the server was handling a
little
MC > over 10MBYTES per secondes, hence very very close to theoretical
limit. At
MC > first I didn’t beleive it > :wink: > So I used somebody else test program.
just to
MC > make sure there wasn’t anything wrong with the way I did the test
which was
MC > basicaly to open ("/dev/null" ) and write() to it. The other test
(based
MC > on Send/Receive/Reply) gave the same results.

MC > It’s the first time I got the see 100Mbits ran that fast. The switch
is
MC > doing a great job but the OS, Net and Net driver must be doing
something
MC > right…

MC > Just though it was worth a post > :wink:

MC > - Mario