(Qos) Save the ANTS, cope with MICE, and shoot the ELEPHANTS

HTB is probably the most commonly root qdisc in use today.

The shapers below it often “shoot” and drop packets in an indiscriminate fashion, paying no attention to traffic type whatsoever, except as configured by the person designing the QoS system.

Shooting the statistically rare yet system critical packets - call them ‘ANTS’ - ARP, DNS, NTP, various routing protocols, DHCP, VPN, and various forms of icmp and (especially) icmp6 has no real effect on the overall load on a router, nor does it reduce bandwidth requirements, and in fact, in many cases (DHCP, VPN, DNS) results in MORE traffic, later, being sent, and often a reduction in user experience that can be quite noticeable (in the case of dns, dhcp, and vpns, in particular. DNS can fall back to TCP….)

Shooting (or, better, marking) a TCP packet instead can reduce bandwidth requirements by huge factors, and there are other protocols where shootdowns (SCTP, etc), has similarly multiplied effects.

Also, TCP MICE are hard to shoot right and have a desirable effect.

Here are some traffic analyses, which can give you a feel for the ratios in one (atypical) home network.

So most shapers should make mildly greater efforts to avoid shooting system critical packets, choosing another packet whenever possible. Simply adding a two or three try mechanism for shoot-downs would nearly eliminate random loss of these types of packets. On the second (or third) try, shoot to kill, however, to avoid gaming the mechanism.

While correct usage of HTB would reduce this problem, by using multiple buckets for multiply classified targets, few classifiers/shapers in the field (try hard enough) (or have the tools to take better aim), so implementing better policy in the kernel would immediately improve the quality of userspace and linux networking.

Similarly, being able to clearly distinguish between Ants, mice and elephants is hard…

Ants

Basically ! tcp suffices as an early match, except when it doesn’t. Further classification gets complex, rapidly.

TCP Mice

The existing conntrack mechanism has the interesting ability to measure the length of a TCP flow. It however, does not do a weighted average or some other ‘mouse detection’ feature like packets/quantum, so as a TCP flow that speeds up and down, cannot migrate from mouse to elephant and back again. Conntrack can only go from mouse to elephant and stay there, which is inappropriate for many protocols such as nfs/cifs filesharing, ssh, etc. That said, it’s soooo close, and effective classification of mice to elephants would aid many shapers, several of which do try to use conntrack in this way.

A common hack in the field (openwrt, at least) is to rate limit syns, (often at very low values (25-50/sec are the openwrt defaults. Syn/ack pairs could be used as a predictor of future work load, or syns could be matched to fins.

Coping with the mice is a hard problem.

Google Test - a,b,c,d,e

A very revealing test is what I call the google ‘a,b,c,d,e’ test - turn on ‘google interfactive’, start capturing packets, and hit a - wait, backspace b…. wait… c… d… e…

And look at the ants, mice, and elephants. Note especially the ratios between successful synacks, the overall length of streams, the start times, the delays, and the other packet losses.

Elephants

Once you can clearly distinguish between ants, mice and elephants, traffic shaping becomes easy. Shooting one elephant can save thousands of other packets… once you choose the right elephant from the herd.

Enhancing ECN

Furthermore, ECN could be used in more cases by more shapers.

so adding a function, drop_or_mark_packet(skb, flag), where flag has the range 0 = SHOOT TO KILL, and 1..X where this marks what to not kill, would be a boon. Return values would be -1 error, 0 if dropped, 1 if marked, 2 if already marked 3..X if one of the reserved packet types.

Psuedocode for the logic would be:

if (v = drop_or_mark_packet(skb,0xFFFFFFFF) > 1) 
{ 
  skb = choose_another_packet(somehow);
  if ( v = drop_or_mark_packet(skb,0xFFFF) > 1) {
     skb = choose_another_packet(somehow);
     v = drop_or_mark_packet(skb,0);
  }    
}

switch(v) {
    -1) Do something about the error;
     0) Log as dropped;
     1) Log as marked; /* Adding a comprehensive 'marking' 
                          structure to the existing 'drop' 
                          statistics is beyond the scope of 
                          this proposal, so you can log as 
                          dropped for now */
}

drop_or_mark_packet would make some attempts (matching against port number and protocol type) to not shoot the aforementioned packet types, except in the case where ECN marking can be used. It might be good to have a more flexible means of controlling this behavior rather than encoding this policy directly in the kernel, as no matter how good the default, the real world is more complex.

and again - this would merely try harder to not shoot the system control packets, ants and mice, but it will if it has to.

choose_another_packet(somehow) may be a non-trivial exercise for some shapers.

Coping with TCP mice is a long difficult subject that needs to be dealt with in some other way, except that shooting fins and fin/acks seems rather counterproductive….

To edit this page, submit a pull request to the Github repository.
RSS feed

Recent News & Articles

Mar 12, 2017 Wiki page
Make-Wifi-Fast Project
Mar 10, 2017 Wiki page
What Can I Do About Bufferbloat?
Feb 10, 2017 Wiki page
More about Bufferbloat
Dec 27, 2016 Wiki page
RRUL Chart Explanation

Find us elsewhere

Bufferbloat Mailing Lists
#bufferbloat on Twitter
Google+ group
Archived Bufferbloat pages from the Wayback Machine

Sponsors

Comcast Research Innovation Fund
Nlnet Foundation
Shuttleworth Foundation
GoFundMe

Bufferbloat Related Projects

Congestion Control Blog
Lede Project (OpenWrt)
Flent Network Test Suite
Sqm-Scripts
The Cake shaper
AQMs in BSD
IETF AQM WG

Network Performance Related Resources


Jim Gettys' Blog - The chairman of the Fjord
Toke's Blog - Karlstad University's work on bloat
Voip Users Conference - Weekly Videoconference mostly about voip
Candelatech - A wifi testing company that "gets it".