mirror of
git://projects.qi-hardware.com/openwrt-xburst.git
synced 2024-12-24 06:45:31 +02:00
691f0d412e
SFQ with external classifiers method. It also corrects a bug in the unsupported ESFQ method already used by qos-scripts. (ESFQ: http://fatooh.org/esfq-2.6/ only updated to 2.6.24, it was switched to an SFQ patch after that and not updated since 2008) A class can be forced to use SFQ, and an external classifier added like this: config class "Normal" option avgrate 10 option priority 30 option packetdelay 100 option limitrate 94 # option qdisc "sfq perturb 2" config class "Normal_up" # option filter "protocol all flow hash keys src divisor 1024" config class "Normal_down" # option filter "protocol all flow hash keys dst divisor 1024" Using these options, the user needs to load cls_flow before qos-scripts starts. I've got more information here: http://oneitguy.com/blogs/netprince/fair-traffic-sharing-esfq-broken-switching-sfqexternal-classifiers This has been tested on r23914. Signed-off-by: Ben Pfountz <netprince<>vt_edu> git-svn-id: svn://svn.openwrt.org/openwrt/trunk@26622 3c298f89-4303-0410-b956-a3cf2f4a3e73
134 lines
3.3 KiB
Awk
134 lines
3.3 KiB
Awk
BEGIN {
|
|
dmax=100
|
|
if (!(linespeed > 0)) linespeed = 128
|
|
FS=":"
|
|
n = 0
|
|
}
|
|
|
|
($1 != "") {
|
|
n++
|
|
class[n] = $1
|
|
prio[n] = $2
|
|
avgrate[n] = ($3 * linespeed / 100)
|
|
pktsize[n] = $4
|
|
delay[n] = $5
|
|
maxrate[n] = ($6 * linespeed / 100)
|
|
qdisc[n] = $7
|
|
filter[n] = $8
|
|
}
|
|
|
|
END {
|
|
allocated = 0
|
|
maxdelay = 0
|
|
|
|
for (i = 1; i <= n; i++) {
|
|
# set defaults
|
|
if (!(pktsize[i] > 0)) pktsize[i] = 1500
|
|
if (!(prio[i] > 0)) prio[i] = 1
|
|
|
|
allocated += avgrate[i]
|
|
sum_prio += prio[i]
|
|
if ((avgrate[i] > 0) && !(delay[i] > 0)) {
|
|
sum_rtprio += prio[i]
|
|
}
|
|
}
|
|
|
|
# allocation of m1 in rt classes:
|
|
# sum(d * m1) must not exceed dmax * (linespeed - allocated)
|
|
dmax = 0
|
|
for (i = 1; i <= n; i++) {
|
|
if (avgrate[i] > 0) {
|
|
rtm2[i] = avgrate[i]
|
|
if (delay[i] > 0) {
|
|
d[i] = delay[i]
|
|
} else {
|
|
d[i] = 2 * pktsize[i] * 1000 / (linespeed * 1024)
|
|
if (d[i] > dmax) dmax = d[i]
|
|
}
|
|
}
|
|
}
|
|
|
|
ds_avail = dmax * (linespeed - allocated)
|
|
for (i = 1; i <= n; i++) {
|
|
lsm1[i] = 0
|
|
rtm1[i] = 0
|
|
lsm2[i] = linespeed * prio[i] / sum_prio
|
|
if ((avgrate[i] > 0) && (d[i] > 0)) {
|
|
if (!(delay[i] > 0)) {
|
|
ds = ds_avail * prio[i] / sum_rtprio
|
|
ds_avail -= ds
|
|
rtm1[i] = rtm2[i] + ds/d[i]
|
|
}
|
|
lsm1[i] = rtm1[i]
|
|
}
|
|
else {
|
|
d[i] = 0
|
|
}
|
|
}
|
|
|
|
# main qdisc
|
|
for (i = 1; i <= n; i++) {
|
|
printf "tc class add dev "device" parent 1:1 classid 1:"class[i]"0 hfsc"
|
|
if (rtm1[i] > 0) {
|
|
printf " rt m1 " int(rtm1[i]) "kbit d " int(d[i] * 1000) "us m2 " int(rtm2[i])"kbit"
|
|
}
|
|
printf " ls m1 " int(lsm1[i]) "kbit d " int(d[i] * 1000) "us m2 " int(lsm2[i]) "kbit"
|
|
print " ul rate " int(maxrate[i]) "kbit"
|
|
}
|
|
|
|
# leaf qdisc
|
|
avpkt = 1200
|
|
for (i = 1; i <= n; i++) {
|
|
printf "tc qdisc add dev "device" parent 1:"class[i]"0 handle "class[i]"00: "
|
|
|
|
# RED parameters - also used to determine the queue length for sfq
|
|
# calculate min value. for links <= 256 kbit, we use 1500 bytes
|
|
# use 50 ms queue length as min threshold for faster links
|
|
# max threshold is fixed to 3*min
|
|
base_pkt=3000
|
|
base_rate=256
|
|
min_lat=50
|
|
if (maxrate[i] <= base_rate) min = base_pkt
|
|
else min = int(maxrate[i] * 1024 / 8 * 0.05)
|
|
max = 3 * min
|
|
limit = (min + max) * 3
|
|
|
|
if (qdisc[i] != "") {
|
|
# user specified qdisc
|
|
print qdisc[i] " limit " limit
|
|
} else if (rtm1[i] > 0) {
|
|
# rt class - use sfq
|
|
print "sfq perturb 2 limit " limit
|
|
} else {
|
|
# non-rt class - use RED
|
|
|
|
avpkt = pktsize[i]
|
|
# don't use avpkt values less than 500 bytes
|
|
if (avpkt < 500) avpkt = 500
|
|
# if avpkt is too close to min, scale down avpkt to allow proper bursting
|
|
if (avpkt > min * 0.70) avpkt *= 0.70
|
|
|
|
|
|
# according to http://www.cs.unc.edu/~jeffay/papers/IEEE-ToN-01.pdf a drop
|
|
# probability somewhere between 0.1 and 0.2 should be a good tradeoff
|
|
# between link utilization and response time (0.1: response; 0.2: utilization)
|
|
prob="0.12"
|
|
|
|
rburst=int((2*min + max) / (3 * avpkt))
|
|
if (rburst < 2) rburst = 2
|
|
print "red min " min " max " max " burst " rburst " avpkt " avpkt " limit " limit " probability " prob " ecn"
|
|
}
|
|
}
|
|
|
|
# filter rule
|
|
for (i = 1; i <= n; i++) {
|
|
print "tc filter add dev "device" parent 1: prio "class[i]" protocol ip handle "class[i]" fw flowid 1:"class[i] "0"
|
|
filterc=1
|
|
if (filter[i] != "") {
|
|
print " tc filter add dev "device" parent "class[i]"00: handle "filterc"0 "filter[i]
|
|
filterc=filterc+1
|
|
}
|
|
}
|
|
}
|
|
|