LibreQoS/sim
2022-12-11 09:05:09 -08:00
..
img Jitter plot 2022-12-10 21:37:23 -08:00
README.org More test tools 2022-12-11 09:05:09 -08:00

Overview

Setting up 4 simulation environments is in order:

Plan Size Characteristics
Small 1024 users
Medium 10000 users
Large 40000 users
Humongous 100k users

IP address ranges

2001:0002::/48 and 198.18.0.0/15 are the experimental address ranges. For testing outside of the network, perhaps a BGP AS for this would help

DNS service

nsupdate for reverse dns? Hosts file?

Subscriber Plans

Plan Type Down Up MinDown MinUP Cap
OldDSL 4Mbit 384kbit 1Mbit
OldDSL6 6Mbit 1Mbit 3Mbit
Cable20 20Mbit 4Mbit 12mbit
LifeLine 25Mbit 3Mbit
FCCmin 25Mbit 3Mbit
NTIAserved 100MBit 20Mbit 20Mbit 5Mbit
Biz50 50MBit 20Mbit 50Mbit 20Mbit
Biz100 100MBit 100Mbit 100Mbit 100Mbit
Biz200 200Mbit 20Mbit 200Mbit 20Mbit
Biz500 500Mbit 50Mbit
Sym1000 1Gbit 1Gbit
Sym100 100Mbit 100Mbit
Sym10 10Mbit 10Mbit

Topologies

A variety of topologies need to be emulated.

Distance

Hops

Different technologies

Routing Protocols

Both ISIS and OSPF are in use

Underlying framing

PPPoe, MPLS are in use CGNATs might become an issue IPv6 deploying?

Tries

Some customers (particularly business ones) have multiple IP addresses. Everyone, using IPv6, has multiple IP addresses.

Worse, these are often dynamic in origin.

Transport Technologies

Plan Down Up
WiFi
Cable 2ms 6ms
Docsis LL 1ms
Wireless half duplex

Test tools

Flent

TRex

goresponsiveness

crusader

speedtest.net

Web PLT

Alexa top 10000

samknows

Netflix video quality

VOIP MOS

Videoconferencing Quality

Congestion Controls

Control Why ecn
cubic Linux S
reno BSD S
ledbat Torrent N
ledbat++ Microsoft Y
curved Microsoft S
bbr Google N
bbrv2 Google exp L
prague IETF exp L

Diffserv Markings

EF,LE,CS1, CS5,NQB,CS6

Failure Modes

AQM

fq_codel was not designed to be a middlebox. The FQ component works well, the AQM component, doesn't work as well as it could.

/IntenseWebs/LibreQoS/media/commit/4aaa0ce90afda8185adab7c07b2abe2a448b11b4/sim/img/middlebox_compared.png

Putting cake on the CPE also results in a metric ton less jitter.

/IntenseWebs/LibreQoS/media/commit/4aaa0ce90afda8185adab7c07b2abe2a448b11b4/sim/img/tonlessjitter.png

Similarly, many devices are themselves the bottleneck, still, so they accumulate a ton of RTT themselves, and monitoring the RTT and doing something about it would possibly help.

Validating the the RTT metrics reported by pping line up with the actual measurements from actual flows is important. Also, what are the effects of ack-filtering on pping?

FQ

The FQ methods we use are really good for most traffic types, and could be even better if more applications did single packet pacing and were more sensitive to delay and jitter.

Encapsulations

We have no insight into QUIC or VPN traffic. This is going to get worse over time. The only thing we have for quic is the: spin bit - which is probably not widely implemented. The best insight we actually have is queue accomulation and packet drop/mark behaviors.

HTB is bursty

DSCP

What DSCPs are in common use today?

DROP_MONITOR

We have 2600 (not kidding, 2600) places where packets can be dropped in the kernel, not to mention other parts of the network. A nice feature would be to be able to track retransmits relative to drops.