From: Barney Wolff <barney@databus.com> To: vjs@sgi.com (Vernon Schryver), nanog@merit.edu
At R=100 SYNs/sec, RTT=250, and L=382, ((L-1)/L)^(RTT*R) = 93%, which is not so bad. Drop-oldest is better with those three numbers, since it works 100% (modulo ordinary problems), but its performance falls off the cliff to 0% at R=L/RTT. If you have a short queue and care about long RTT's, random drop is better than drop-oldest.
Agreed. Note that 93% is not bad for a human-initiated telnet, but is disastrous for a Web browser which initiates a dozen tcp sessions to retrieve one page, because the browser will probably not retry at all if it gets a reset, but instead report failure to retrieve the page to the user, who can only ask it to start over from the beginning.
Caching should keep them from having to retransfer everthing. Of course, if the problem is getting through the listen queue, that might not help much.
So I think that it's better to accept the limited-radius-under-attack property of drop-oldest, gaining the immunity from interference within the safe radius.
If you have a useful safe radius. At only 130 SYNs/sec, a 260 entry queue seems useful today, giving a 2 second radius. What is the limit on the queue in the typical router or Ethernet hub?
If it were possible to set the syn-rcvd timeout with sub-second granularity, this "fix" would not even take any kernel code mods - but of course it does not adjust the safe radius dynamically as the attack rate changes.
Again, I think it is practical to switch automagically from drop-oldest when the attack rate is modest to random-drop when the bad guys get something better than a v.34 modem, such as ISDN or just a corrupted machine somewhere on an Ethernet connected with a T1.
What's absolutely clear is that any method of queue pruning is better than none, and a big queue is required for survival.
A queue limit 50 to 500 times larger than the 7 from the classic BSD SO_MAXCONN=5 seems useful for handling the 100's of valid TCP connections/sec that you get on a busy WWW server. Vernon Schryver, vjs@sgi.com