This document is meant as quick braindump. Take it as such.
Blackhole efect of a victim
Victim of DDoS is acting (and being treated by attacker) as blackhole
for IP packets. Many packets are flowing in and a little reply comes
out of it. It hold as long as we consider duplex streams only. As
multicast streams are not allowed for regular users the assumption
should be generaly true.
And we can use it to detect a victim.
Victim detection at leaf gateway
Leaf gateway (where no multipath is in use toward client systems) should
do very simple connection tracking of IP it routes. When it routes packet
into backbone it creates record for src-dst IP addresses. It allows sending
packets for some time T (say 30 secs) without actually geting a reply packet.
Once reply packet has been seen the connection is "confirmed" for the
same time T.
If no reply packet is recieved then forward packets are shaped to rate
1 packet per T until a reply packet is recieved.
The result is that if this would be implemented on as many as possible leaf
gateways (or even on larger backbone systems) it would allow victim to
mitigate an attack by simply stopping its server and begin to gradualy
start service for increasing parts of client IP space.
Performance of implementation
When implementing this on Gbps link then we need to look whether it is possible
with moderate ammount of HW. One conntrack record would need to carry at least
- state - 1 byte
- timestamp - 3 bytes
- IP pair - 8 bytes
Thus we need 8-12 byte records (depending on data structure) per conversation.
Probably the best is to use stochastic approach and create hash table with size
of 2^24 records which would need 160MByte of dedicated memory. The table would
be keyed by last 3 octets of destination IP and acting as 1-way cache.
This way some records are not in table and can go thru. But we need not to kill
100% of data, 99% is good too. This approach should work well for systems
with up to tens of thousands client IPs.
This filter can be implemented by fast PC runing some *nix system or dedicated
hardware. One can take 2 pieces of 1Gbit chipset, FPGA in $100 range and implement
the filter as 3 stage pipeline tied by dualported mems (fetch info for a packet,
decide packet's fate, act on the packet and write the info back).
With latest Spartan3 from Xilinx I can imagine cheap design with filter rate
10 Mpackets/s (use 133MHz/50ns SDRAM) which would be able to sustain 1Gbit rate.
Not bad for under $200 design.
Calculating safety measure
When implemented by majority of Internet with T=30s and number of attacking
machines N=100.000 we get rate r=1500*N/T=50MBit. It is better than approx
5Gbit it could be without this system but still not so good.
One could exponentially increase T of records which still sends packets
against blackhole. T could reach something like 1hr which would allow victim
to mitigate the attach under 1MBit which is likely manageable.
Possible dangers
If implemented in system where IP spoofing is possible then one could "install"
conntrack info for other clients. So ISP implementing this must implement source
route validation too.
Another benefits
The system would allow any host to develop/buy more advanced DDoS detector
while blackhole detectors are providing framework doing actual filtering.
No cross-administrative-domain cooperation needed.
Port scanning made much more complex.
devik placeathere cdi placedothere cz