[ale] 100 vs 10 Mbit problem

[Dow Hurst]

Thanks all for the advice!  The Bay Networks switch is older but managed
so I used it.  I can replace it with a newer Linksys 16 port switch as a
test.  Also, I can have a 5 port switch I can stick for another test. 
The memory issue could be the reason since both 10Mbit machines show the
collisions occuring.  I didn't see this on a 8 port DLink switch with
the same machines.

Just goes to show how technology moves on.  The Baynetworks switch was
$4K in 1997 and the DLink was $79.00 in 2001.
Dow

PS. The machines with the 10Mbit interfaces are an old R8000 Power
Indigo2 and an even older 4xR10K Onyx from 1993.  In fact, the SGI Onyx
is the same one that was given away on the ALE list a couple of years
ago.  We had it refurbished and it is running jobs all the time for us. 
Not the fastest machine but it works.

>>> jonathan at xcorps.net 09/12/03 21:20 PM >>>
On Friday 12 September 2003 18:51, Dow Hurst wrote:
> I have two machines with 10Mbit interfaces on a Baynetworks managed
> 24port 10/100Mbit switch.  The other machines on the switch are
> 100Mbit.  I am seeing collisions on the slower interfaces and having
> trouble with large data transfers to and from those machines.  Is
there
> a solution to this problem that I can implement?  I haven't seen this
> before even though I've had slow and fast ethernet interfaces on
> negotiating switches before.  Should I change cables first, and try
> different ports before anything else?

This actually was a very common problem back in the mid to late 90s when

100baseT first started springing up. It is known as switch overload. The

problem is most likely the switch itself. Someone else mentioned adding
a 
second switch. If that works then it is definitely a case of switch 
overload. What's happening is, you've got a firehose and a drinking 
straw. The switch is a large funnel. The switch's memory pool is the 
large conical part of the funnel and it can hold one gallon of water.
The 
switch's memory gets full and it starts discarding packets. Adding the 
second switch adds another memory pool. It sounds to me like you have a 
duplexing mismatch involved here as well. Stepping the 100Mb machines 
down to HD will help a lot, but there is no ideal way to make this 
problem go away. The choices are, slow everything down to the speed of 
the slow machines, upgrade the switch fabric either by adding a second 
switch or upgrading the existing switch, or put in a hub or slower
switch 
until you can get faster NICs for the slow machines. If you can manage 
the switch on a per-port basis...set all ports to HD. When I ran into 
this problem on a large scale, I ended up reverting to 10Mb switches in 
several places just because visiting each workstation was not possible
on 
short notice. Balance is very important on switched networks. 

-- 
Jonathan Rickman
Key ID: 0DF501FF


_______________________________________________
Ale mailing list
Ale at ale.org
http://www.ale.org/mailman/listinfo/ale