[Bloat] industrial scale bufferbloat in a DDOS on core DNS servers

[Dave Taht]

I had a lot of papers to read on my vacation. This one, in section 3.3:

"Figure 7 shows the median RTT for selected K-Root sites. Although the
K-AMS site remained up and showed minimal loss, its median RTT showed
a huge increase: from roughly 30 ms to 1 s on Nov. 30, and to almost 2
s on Dec. 1, strongly suggesting the site was overloaded. K-NRT shows
similar behavior, with median RTT rising from 80 ms to 1 s and 1.7 s
in the two events. Overload does not always result in large latencies.
B-Root (a single site) showed only modest RTT increases (Figure 4),
since only few probes could reach it during the attack (Figure 3). We
hypothesize that large RTT increases are the result of an overloaded
link combined with large buffering at routers (industrial-scale
bufferbloat [23])."

from: http://www.isi.edu/~johnh/PAPERS/Moura16a.pdf

The authors did not have any insight into where on the path the RTT
increases were coming from. Now that we have adequate fq and aqm
solutions out there for bsd and linux, perhaps some load balancers
(often bsd based?) are sources of bufferbloat? Or perhaps others in
the dns world, fighting off DDOS attacks, can look harder at where the
RTT increases are from?


-- 
Dave Täht
Let's go make home routers and wifi faster! With better software!
http://blog.cerowrt.org