Network Operators and smurf
CMartin at mercury.balink.com
Sun Apr 26 06:08:04 UTC 1998
Given the contributions to this thread have been mostly theoretical in
nature, I'd like to share an experience of mine that in some ways
negates some of the propsed solutions to smurf attacks in the context of
Recently, one of our downstream customers was subject to a smurf attack,
and we placed an access-list on our egress interface to the customers
network. The customer hangs of an SMDS cloud - our link to the cloud is
at 34 Mbps, his T1. We were blocking echo replies destined for his
network. We are connected upstream at 45Mbps. As the attack
intensified, router CPU Utilization jumped to 99%, and the input queue
on our inbound HSSI was at 75/75. We started dropping packets at a rate
of about 7000/sec. The attacks were coming in from all over the world.
The NetFlow cache was growing at an alarming rate, and, after a while
the HSSI just DIED. As the HSSI bounced, our BGP session bounced with
it, causing some mild route flapping (Not vaccilating enough to be
damped, but enough nonetheless). Eventually the attack subsided, and
all went back to normal, but for a period of time, say 10 minutes, we
had a 7507, 64megs, RSP2 WITH the HSSI on a VIP2 on its knees.
We decided that parsing NetFlow logs would give us a better idea of who
was amplifying the attacks, and with a simple shell script, we were able
to build a database of ASNs, with admin contacts from RADB/RIPE/ARIN.
We are planning on sending emails to these customers to ask them to stop
amplifying smurfs (script does that too), because this is unacceptable.
My point, then is this: Filtering echo replies is/can be a futile
attempt at preventing this type of attack. I watched a 7507 die
defending against one attack. More recently, we got hit so hard that
the router was screaming without ANY access-lists blocking ICMP
echo-replies, perhaps because there was no real fast-switching taking
place (each source is different, so the first packet is process
switched. Our NetFlow cache went from 3900 flows to 27000 flows in
about 4 mins.) And, as we were not amplifying the smurfs, source
address verification is a moot point. I am all for allowing these
netblocks time to implement this type of filtering (layer3 to layer2
broadcast translation prevention), but not for very long. It appears
the best way to light a fire under someones rear end is to publicly
shame them into acting. For those who don't act quickly enough, there
can be no quarter.
If I appear hostile, I am...
If anyone has similar experiences, please share them, as there has
already been enough rhetoric filling this thread, and it is clear that
everyone knows the solution lies at edge and beyond, not in the core.
More information about the NANOG