[6524] in North American Network Operators' Group
Re: NAP/ISP Saturation WAS: Re: Exchanges that matter...
daemon@ATHENA.MIT.EDU (Vadim Antonov)
Tue Dec 17 00:07:17 1996
Date: Mon, 16 Dec 1996 20:58:47 -0800
From: Vadim Antonov <avg@pluris.com>
To: cnordin@vni.net, doleary@cisco.com
Cc: nanog@merit.edu
Craig Nordin <cnordin@vni.net> wrote:
>Interesting. An ICMP packet dropped when busy. Well, it seems as
>if there is only a hair's difference between when an ICMP packet is
>dropped and when an IP packet is dropped.
Well, cisco (and all other known fast routers, for that matter)
handles ICMP in software, unlike vanilla IP packets, which
are routed with significant hardware assistance.
I.e. a lot of ICMPs will overload CPU. It won't interfere
with routing of IP packets (though it can wreck havoc with of
routing and network management protocols). So cisco limits
the rate at which ICMPs are processed.
>If you are busy, you are busy, right?
Not. Different pieces of hardware are busy in those cases.
>I know that I was getting zero packet loss for many many basic routes
>this time last year that are now losing packets. I think that a network
>is in great shape when the packet loss is at a sheer minimum. Even one
>percent packet loss can be felt as substantially more degraded than
>perfect transport.
Er. There's no such thing as perfect transport as long as TCP is
concerned. If end-hosts support large windows, even a single
TCP session will load the network to the point where it'll lose
packets.
>Just like ra.net, I use pings to monitor one aspect of overall performance.
>Me and ra.net are not alone.
Pings are ok, as long as you realize what exactly is being measured.
If you're pinging ciscos, you aren't measuring IP packet loss.
--vadim
