nanog mailing list archives

Re: Any 1U - 2U Ethernet switches that can handle 4K VLANs?


From: Will Hargrave <will () harg net>
Date: Mon, 26 Jan 2004 00:18:27 +0000


On Sun, Jan 25, 2004 at 09:39:05PM +0100, Mikael Abrahamsson wrote:
This is interesting, what problems did you run into?

We have an extensive Extreme networks used both for L2 and L3, and apart 
from the fact that it always cpu routes ICMP, I see no major flaw in the 
L3 forwarding function (for access/distribution) for all normal purposes.

ACLs are per-port and known to be buggy when operating on port numbers -
in particular UDP ACLs match will match arbritary data when presented
with a subsequent IP fragments (think NFS...)

As pointed out in a similar thread recently, the 'flow-based' (well, 
destination IP based) ipfdb will crap out on the Extremes under heavy load 
- e.g. virus'd machines internal to your network doing heavy scanning.
Symptom is very poor performance and the 'top' command will show heavy
CPU usage as subsequent flows are CPU routed.

My few experiences with the Cisco 3550 as L3 routers has been much worse, 
even with claimed CEF capability I have seen it melt and die where the 
equivalent Extreme box didnt experience the same problems (of course there 
are cases where it's the other way around). Overall I have more confidence 
in the Extreme access boxes for L3 than Ciscos equivlanent, and they 
definately kick ciscos ass when it comes to L2 (mac address table size and 
number of vlans for instance).

The 'recommended max' number of SVIs for the 3550 is something low like 8.
There is no limited stated in the datasheet for the 3750 - is anyone
running more than 8 SVIs on a 3750? 

The ACL capability on the 3550 seems a lot more capable but the lack of
unicast RPF is irritating. (More irritating, 'ip verify unicast
reachable-via...'  is accepted but silently does nothing)

I'd be very interested to hear what conditions you've found cause
problems for Cat3550s. We're planning to buy quite a few more of this range
(probably 3750-24) to reduce L2 size in our network and for CPE-type
uses. 



Current thread: