1. Detecting ICMP Rate-Limiting Les Cottrell Warren Matthews Mit Shah

2. Motivation 1 n Smurf Attacks : IP spoofing and ICMP packets to an IP Broadcast group. Traffic to target multiplied by responses from each member n Example : Attacker on 768Kbps stream and a 100 member Broadcast group generate 77Mbps of traffic and swamp target! n Routers set to : “no ip directed- broadcast”

3. Motivation 2 n Cisco introduces CAR (Committed Access Rate) in 7200 and 7500 series routers. Later includes support in IOS 12.0 n access-list 102 permit icmp any any echo n interface Serial3/0/0 rate-limit input access-group 102 256000 8000 8000 conform-action transmit exceed-option drop

4. ICMP Blocking - No Response! n www.vincy.bg.ac.yu blocked 884 rounds of 10 ICMP packets each, out of 903 n islamabad-server2.comsats.net.pk blocked 554 out of 903 rounds n leonis.nus.edu.sg blocked all packets it was sent (All examples from data for Dec 1999) n Yet in reality, none of these servers was down!

5. New tools to the rescue n SYNACK developed in-house n Establishes TCP connections and measures time taken by target to respond n Cleans up connections n Highly visible to system admins n STING developed by Stefan Savage n TCP can’t ack out- of-order packets n Data-seeding and Hole-filling (reliable) n Need to change one line of kernel code

6. Results from Sting & Synack n Both tools based on TCP/IP, hence appear to router to be “normal” traffic n Results : n The Singapore node responds ONLY to 56+8 byte packets n Both the other nodes were alive-and- kicking with low loss rates!

7. Utility of Sting as an aid n These are 5 sites that were responding to pings very infrequently, and neglecting entire sets of 10 pings more than 50% of the time n Sting showed that they were alive on port 80!

8. Tail-Drop Behavior n Rate-limiting kicks in after the first few packets and hence later packets are more likely to be dropped n This node no longer displays tail-drop behavior!

9. Frequency Analysis n Calculate the packet drops as a function of packet-numbers n Calculate the slope and identify extremes n Implemented by Warren as a metric n Some encouraging early results!

10. Some Candidates :

11. CAR (Committed Access Rate) n Tokens removed in proportion to size of packet n Maximum number of tokens in bucket = Normal Burst Size n Extended Burst mechanism to make drops more RED- like

12. RED (Random Early Detection) n Tail-drop causes packet-loss across all TCP streams when traffic is too heavy n Causes all TCP-streams to sense congestion and start recovery n Small, bursty TCP streams also have to restart n Solution : drop packets randomly BEFORE congestion strikes!

13. Extended Burst Mechanism in CAR n Stream allowed to borrow more tokens if extended-burst value > normal-burst n “ Compounded debt” computed as sum of a(j) where j denotes the jth packet that tries to borrow tokens since last packet drop and a(j) denotes actual debt value n Packet dropped if CD > extended-burst and CD set to 0

14. Detecting CAR : the good news n A stream at constant rate R, above the configured-rate C, will exhaust tokens in bucket after B/(R-C) sec, at most n From this point on, borrowed packets at jth packet = j*(R-C) and beyond j=E/(R- C), actual debt > extended-burst and all packets will be dropped n Pattern is non-random!

15. Detecting CAR : A trial n Analyzed the first-order differences in packet-numbers of dropped packets to see if there was a pattern hoping that site-specific CAR might have set packet-size > normal_burst_size + extended_burst_size n Not surprisingly, no results n False alarm : 10th packet being dropped but data was TOO clean!

16. Detecting CAR : the bad news n It appears that most sites will impose a traffic-limit on TOTAL icmp traffic n Predicting when a packet drop occurs akin to predicting the rest of the traffic on that router at that moment - a known “hard” problem! n Solution : Aggressive pinging, your traffic-stream dominates! High signal-to- noise!!!

17. Further study n Pinging with variable-sized packets (less than MTU) and detect whether packet-loss varies linearly with size n trivial to determine MTU? n How important are other effects like being more likely to be dropped from queue? n Set up a router that implements CAR, simulate icmp traffic, and study patterns