100mbps switched network, slow transfer rates.

Hi,

I am about to cry, I cannot seem to figure this one out. I have several
boxes running various different builds of Linux. All connected with
some random Cisco (not at farm, not entirely sure what model) 100Mbps
switches.

On an idle network one would expect a nice 10MB/Sec transfer rate
between machines, alas this is not the case. Some boxes can do this,
some can only download at this, and others just give up the ghost.
There is no difference between the boxes, they are all running the same
hardware.

Now I have tried different network cards etc, but the results are the
same. I have two boxes which are both running SuSE 9 (one with
2.6.0-test5-10-smp and the other with 2.4.21-144-smp4G). Plus other
boxes running fun things like RedHat - and same results.

All boxes have 100Mbps Full-Duplex forced, and the switch seems to
confirm this. Although the through-put is indicating 10Mbps Half-Duplex
at the best. I cannot for the life of me figure out why this is. I've
checked hard drive I/O, which is all normal, the systems are all idling.

Is there anybody who can help, hint, or suggest anything that might be
able to help me?

I am using these nics:
3Com 3C905C-TX Fast Etherlink
Intel EtherExpress PRO/100+

Cheers,

Alastair.

--
http://www.wireplay.co.uk/
http://www.aligrant.com/

"Never be ashamed of not knowing a
thing, but always be ashamed of not
wanting to know."

> I am about to cry, I cannot seem to figure this one out. I have several
> boxes running various different builds of Linux. All connected with
> some random Cisco (not at farm, not entirely sure what model) 100Mbps
> switches.
>
> On an idle network one would expect a nice 10MB/Sec transfer rate
> between machines, alas this is not the case. Some boxes can do this,
> some can only download at this, and others just give up the ghost.
> There is no difference between the boxes, they are all running the same
> hardware.

(snip)

> Is there anybody who can help, hint, or suggest anything that might be
> able to help me?

This type of thing comes up incredibly often. I'll suggest that you make
use of the quite voluminous information that's already been given, by using
http://groups.google.com to see what's already been covered.

One other thing that jumps out is that the switches are Ciscos. I've
never seen *any* products with worse autonegotiation than Cisco. Even with
two Cisco devices connected directly with full-spec, non-defective cable,
both having "negotiated" the same speed, I've seen weird, bizarre crap like
what you're seeing until both have the speed and duplex settings set
manually. Try manually setting the ports in question to 100/full, and see
if that makes a difference. That particular bugginess bit a LOT of people
in the butt when the big Cisco DOS exploit came out some time ago - everyone
in the world was upgrading their IOS, which is notorious for not always
preserving all of your settings. Suddenly, routers all over the Internet
were back to autonegotiation, without their admins realizing it. I had a
good number of customers complaining of various problems, and every one of
them was traced to their provider's Cisco equipment being back to
autonegotiation. Alright, that's enough of a rant about Cisco. : )

Also, try compiling your own kernels, with only the networking options you
*need*. I've seen an improvement of several megaBYTES/second by switching
from a RedHat kernel (which usually has every option/module imaginable) to
one compiled with only what I needed.

There's also the questions of what *type* of transfer you're doing, how
powerful your hardware is, if the cabling has been checked/replaced, and the
other "standard responses" that you'll find by searching through google.

steve

Alastair Grant wrote:
> All boxes have 100Mbps Full-Duplex forced, and the switch seems to
> confirm this. Although the through-put is indicating 10Mbps Half-Duplex

If "forced" means that autonegotiate is turned off, the switch must have
it turned off as well; the spec says that if negotiation fails, the
default
of half-duplex should be used.
wfg -job