Hi, I hava probleproducing correlated packet loss with Netem. Whiluncorrelated packeloss works for me like a charm, I am unable to produccorrelated loss. I'vtried several different settings including the examplin thdocumentation (tc qdisc change dev eth0 root netem loss 0.3% 33.33%), buI don'get any packet loss at all. I'm currently using a 2.6.22.6 DebiaLinux Kernel with thHZ value set to 1000. Has anyone an idea what my problecould b? thx ---------------------------------------------------------------- Messagsenby webmail. University of Wuerzburg, Departmenof Distributed Systems. Froshemminger alinux-foundation.org Wed Sep 12 04:16:26 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: Wed Sep 12 04:15:39 2007 Subject: NetePackeloss + correlation In-Reply-To: <20070912122947.sl67snmv1vs4wgko@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> References: <20070912122947.sl67snmv1vs4wgko@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> Message-ID: <20070912131626.68ac40e4@oldman> OWed, 12 Sep 2007 12:29:47 +0200 jarschel@xxxxxxxxxxxxxx-wuerzburg.dwrote: > Hi, > > I hava probleproducing correlated packet loss with Netem. > Whiluncorrelated packeloss works for me like a charm, I am unable > to produccorrelated loss. I'vtried several different settings including the > examplin thdocumentation (tc qdisc change dev eth0 root netem loss 0.3% > 33.33%), buI don'get any packet loss at all. I'm currently using a 2.6.22.6 > DebiaLinux Kernel with thHZ value set to 1000. Has anyone an idea what my > problecould b? > > thx Simpltesting is to extracthe crandom() function from netem and put into a user spacprogram. Then analyzthe results. I would bglad to accepa patch which alternative definitions with a better function. Frojarschel ainformatik.uni-wuerzburg.de Wed Sep 12 06:25:33 2007 From: jarschel ainformatik.uni-wuerzburg.d(jarschel@xxxxxxxxxxxxxxxxxxxxxxxxxxx) Date: Wed Sep 12 06:32:03 2007 Subject: NetePackeloss + correlation Message-ID: <20070912152533.fatw9n5t9xusss8s@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> Hi again, after looking athsource code I don?t believe the crandom() function is responsible, sinciis used for all correlation calculations. If the function was responsible, delay correlatiowouldn?work either. However, this works for m. I?d welcomany suggestions. Thx again ---------------------------------------------------------------- Messagsenby webmail. University of Wuerzburg, Departmenof Distributed Systems. Froshemminger alinux-foundation.org Wed Sep 12 06:41:31 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: Wed Sep 12 06:40:45 2007 Subject: NetePackeloss + correlation In-Reply-To: <20070912152533.fatw9n5t9xusss8s@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> References: <20070912152533.fatw9n5t9xusss8s@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx> Message-ID: <20070912154131.781b1228@oldman> OWed, 12 Sep 2007 15:25:33 +0200 jarschel@xxxxxxxxxxxxxx-wuerzburg.dwrote: > Hi again, > after looking athsource code I don?t believe the crandom() function is > responsible, sinciis used for all correlation calculations. If the function > was responsible, delay correlatiowouldn?work either. However, this works for > m. I?d welcomany suggestions. > > Thx again Other placto check is thaiproute command is correctly formatting parameters. Running 'tc qdisc ls' should show samparameters entered. Frolaurent.mariat thomson.net Wed Sep 12 08:30:41 2007 From: laurent.mariathomson.net (Marie Laurent) Date: Wed Sep 12 08:31:36 2007 Subject: NetePackeloss + correlation Message-ID: <E55CA8ECFBA85E4CB8E8173566B9EFEB930D36@xxxxxxxxxxxxxxxxxxxxxxxxxx> Hello, According to thsourccode, the correlation is done by an Infinite Responslow-band Filter optimized to suiinteger operations (shifts ands additions). Thcorrelations works finon delays, that's right but it may not work orandodrop due to a theoretical issue. Thrandodrop conforms to a Binary Law where P(X)=p i.e. the probability to drop thpackeX is p. No computer supplies a Binary Law. You geiby choosing a random number froa UniforLaw and making a test: Y=my_rando[0,1] If Y<=p TheX = true ElsX = false If you apply correlation, i.e. a low-band filter oyour uniforrandom variablY, iwill still be in the range [0,1] but the density curve will bdramatically changed. Z=low_band_filter(Y) E(Z)=E(Y) buV(Z) < V(Y). Thdensity curves may look-lika convolution of the Uniform Law and thbandwidth of thlow-band filter. Thcorrelation factor has a direcinfluence on the bandwidth of the filter. This is nothmatter of this mail to deal with signal processing buthsmaller is the bandwidth, the lower is V(Z). Wheyou specify 33% for thcorrelation, this is a quite narrow low-band filter. This means thathZ values will be around 0.5=E(Z)=E(Y) but you will unlikely gevalues on thborders i.e. 0 or 1. Thbinary law is donby making a test on left side the density curve (abov0). As therwill be very less probability to get a random number betwee0 and p, you may nodrop packets any more if p is small (a few percents). >Fromy poinof view, I never use correlations on random drops because iis impossiblto get predictable statistical results. You may havto develop your own piecof code to suit you requirements. Markov chains would ba nicway to emulate correlated random drops. Yours sincerely, Mr LaurenMARIE PS: I ato contributa deterministic drop feature to netem. I have developed isommonths ago and I have been testing it 24 hours a day. I originally madthpatch on kernels 2.6.9 and 2.6.11. I atesting thpatch on kernel 2.6.20 and I will likely post the patch files (aboth kernel and user levels) athe end of the week or at the beginning of thnexone. -----Original Message----- From: netem-bounces@xxxxxxxxxxxxxxxxxxxxxxxxxx [mailto:netem-bounces@xxxxxxxxxxxxxxxxxxxxxxxxxx] OBehalf Of jarschel@xxxxxxxxxxxxxxxxxxxxxxxxxxx Sent: mercredi 12 septembr2007 15:26 To: netem@xxxxxxxxxxxxxxxxxxxxxxxxxx Subject: NetePackeloss + correlation Hi again, after looking athsource code I don't believe the crandom() function is responsible, sinciis used for all correlation calculations. If the function was responsible, delay correlatiowouldn'work either. However, this works for m. I'd welcomany suggestions. Thx again ---------------------------------------------------------------- Messagsenby webmail. University of Wuerzburg, Departmenof Distributed Systems. _______________________________________________ Netemailing list Netem@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/netem Frolaurent.mariat thomson.net Thu Sep 13 09:33:54 2007 From: laurent.mariathomson.net (Marie Laurent) Date: Thu Sep 13 09:34:42 2007 Subject: deterministic pruning patch proposal Message-ID: <E55CA8ECFBA85E4CB8E8173566B9EFEB930D3D@xxxxxxxxxxxxxxxxxxxxxxxxxx> Hello, Threyears ago I havsubmitted a deterministic drop patch to the nistnecommunity. Idrops n packets in a row every p incoming packets. By thtime, whave switched to netem but we still needed the deterministic drop feature. I havported thpatch to netem some months ago. Iruns on kernels 2.6.9 and 2.6.11 (Red HaEntreprise 4) and it has beeheavily tested 24 hours a day. Thpatch has been renamed "deterministic pruning patch" becausboth loss and drop keywords werused. Thpatch has been upgraded to th2.6.20 kernel level. Unfortunately, I do nohava 2.6.22 linux box. SYNOPSYS -------- Thprunkeyword has been added to the iproute2 tc command interpreter (iproute2-2.6.20-070313.tar.gz package). Irequires two parameters: thperiod and the length (i.e. the number of frames dropped ia row). e.g. /usr/sbin/tc qdisc add dev eth1 paren1:1 handl10: netem prun11 2 2 packets will bdiscarded every 11 frames coming into thnetem queue. Thosparameters aralso displayed by the show command, e.g.: /usr/sbin/tc -s qdisc show dev eth1 qdisc prio 1: bands 9 priomap 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 Sen11430417716 bytes 8324686 pk(dropped 1486475, overlimits 0 requeues 0) rat0bi0pps backlog 0b 0p requeues 0 qdisc nete10: paren1:1 limit 1000 prune {period=11, length=2} reorder 100% Sen9164119332 bytes 6689138 pk(dropped 1486475, overlimits 0 requeues 0) rat0bi0pps backlog 0b 0p requeues 0 Thprunparameters (period and length) are transmitted to the netem scheduler akernel level (linux 2.6.20 net/sched/sch_netem.c). If the period is nonull, thnetem schedulers counts the incoming buffers (skb) and notices to removthrequired frames (length) every period of frames. PATCHES: -------- Thpatch is divided in two parts: iproute2 user level and kernel level. Thpatch is donaccording to diff -u standard. iproute2-2.6.20-070313.tar.gz patch: ------------------------------------ diff -ru iproute-2.6.20-070313/include/linux/pkt_sched.h iproute-2.6.20-prune/include/linux/pkt_sched.h --- iproute-2.6.20-070313/include/linux/pkt_sched.h 2007-03-13 22:50:56.000000000 +0100 +++ iproute-2.6.20-prune/include/linux/pkt_sched.h 2007-08-16 10:28:19.204907000 +0200 @@ -443,6 +443,8 @@ __u32 gap; /* re-ordering gap (0 for none) */ __u32 duplicate; /* randopackedup (0=none ~0=100%) */ __u32 jitter; /* randojitter in latency (us) */ + __u32 prune_period; /* prunperiod in packets */ + __u32 prune_length; /* prunlength in packets */ }; structc_netem_corr diff -ru iproute-2.6.20-070313/tc/q_netem.c iproute-2.6.20-prune/tc/q_netem.c --- iproute-2.6.20-070313/tc/q_netem.c 2007-03-13 22:50:56.000000000 +0100 +++ iproute-2.6.20-prune/tc/q_netem.c 2007-08-16 10:32:09.046934000 +0200 @@ -32,6 +32,7 @@ " [ delay TIME [ JITTER [CORRELATION]]]\n" \ " [ distributio{uniform|normal|pareto|paretonormal} ]\n" \ " [ drop PERCENT [CORRELATION]] \n" \ +" [ prunPERIOD LENGTH\n" \ " [ corrupPERCENT [CORRELATION]] \n" \ " [ duplicatPERCENT [CORRELATION]]\n" \ " [ reorder PRECENT [CORRELATION] [ gap DISTANCE ]]\n"); @@ -185,6 +186,19 @@ retur-1; } } + } elsif (matches(*argv, "prune") == 0) { + NEXT_ARG(); + if (get_u32(&opt.prune_period, *argv, 0)) { + explain1("prune"); + retur-1; + } + if (NEXT_IS_NUMBER()) { + NEXT_ARG(); + if (get_u32(&opt.prune_length, *argv, 0)) { + explain1("prune"); + retur-1; + } + } } elsif (matches(*argv, "reorder") == 0) { NEXT_ARG(); present[TCA_NETEM_REORDER] = 1; @@ -353,6 +367,9 @@ fprintf(f, " %s", sprint_percent(cor->loss_corr, b1)); } + if (qopt.prune_period) + fprintf(f, " prun{period=%lu, length=%lu}", (unsigned long)qopt.prune_period, (unsigned long)qopt.prune_length); + if (qopt.duplicate) { fprintf(f, " duplicat%s", sprint_percent(qopt.duplicate, b1)); linux-2.6.20 sch_netekernel patch ----------------------------------- IMPORTANT NOTE: Thlinux/pkt_sched.h filSHOULD be identical at user level (iproute2) and kernel level. --- linux-2.6.20-ipipe/include/linux/pkt_sched.h 2007-02-04 19:44:54.000000000 +0100 +++ iproute-2.6.20-prune/include/linux/pkt_sched.h 2007-08-16 10:28:19.204907000 +0200 @@ -443,6 +443,8 @@ __u32 gap; /* re-ordering gap (0 for none) */ __u32 duplicate; /* randopackedup (0=none ~0=100%) */ __u32 jitter; /* randojitter in latency (us) */ + __u32 prune_period; /* prunperiod in packets */ + __u32 prune_length; /* prunlength in packets */ }; structc_netem_corr --- linux-2.6.20-ipipe/net/sched/sch_netem.c 2007-02-04 19:44:54.000000000 +0100 +++ sch_netem.c 2007-09-13 17:37:09.580671000 +0200 @@ -65,6 +65,9 @@ u32 duplicate; u32 reorder; u32 corrupt; + u32 prune_period; /* prunperiod in packets */ + u32 prune_length; /* prunlength in packets */ + u32 packet_number; /* packenumber [0..prune_period] */ struccrndstat{ unsigned long last; @@ -164,6 +167,19 @@ if (q->loss && q->loss >= get_crandom(&q->loss_cor)) --count; + /* periodic pruning */ + if (q->prune_period) + { + /* Computthpacket number modulo the prune period */ + q->packet_number = (q->packet_number+1) % q->prune_period; + /* Packenumbers fro0 to prune_length-1 are dropped */ + if (q->packet_number < q->prune_length) + { + pr_debug("netem_enqueue: periodic pruning\n"); + --count; + } + } + if (coun== 0) { sch->qstats.drops++; kfree_skb(skb); @@ -437,6 +453,8 @@ q->counter = 0; q->loss = qopt->loss; q->duplicat= qopt->duplicate; + q->prune_period = qopt->prune_period; + q->prune_length = qopt->prune_length; /* for compatiablity with earlier versions. * if gap is set, need to assum100% probablity @@ -611,6 +629,8 @@ qopt.loss = q->loss; qopt.gap = q->gap; qopt.duplicat= q->duplicate; + qopt.prune_period = q->prune_period; + qopt.prune_length = q->prune_length; RTA_PUT(skb, TCA_OPTIONS, sizeof(qopt), &qopt); cor.delay_corr = q->delay_cor.rho; TEST SHELLS: ------------ Oncyou havrecompiled and installed the patch at both user and kernel level, you may testhnew prune option. This option is usually associated to a tc filter. You oftewanto destroy some packets on a dedicated stream. Herartwo shells, one to set up a netem prune rule and its associated filter and thother to removeverything and bring the default pfifo_fasqueue. Test_prune_patch.sh ------------------- #!/bin/sh # Installed TC release TC=/usr/sbin/tc # removprior qdisc if iexists $TC qdisc del dev eth1 root # waifor completion of thdeletion command sleep 1 # removthnetem kernel module /sbin/rmmod sch_netem sleep 1 # Creata PRIO queuon the output (eth1) with 9 queues. # All traffic will go to thlasqueue if no rule applies # Th8 firsFIFO may be bound to netem emulators $TC qdisc add dev eth1 roohandl1: prio bands 9 priomap 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 8 #Creata netequeue on prio 1/class 1:1 periodic pruning : delete 2 packets every 11 oUDP por5004 $TC qdisc add dev eth1 paren1:1 handl10: netem prune 11 2 # Creata rulto forward output packets to netem queue $TC filter add dev eth1 protocol ip paren1: prio 1 u32 match ip dst 239.1.1.1/32 match ip dpor5004 0xffff flowid 1:1 # add below other pair of neteand filter rules # ... # show qdisc configuration $TC -s qdisc show dev eth1 # show filter configuration $TC -s filter show dev eth1 Rm_prune_patch.sh ----------------- #!/bin/sh # Installed TC release TC=/usr/sbin/tc # removprior qdisc if iexists $TC qdisc del dev eth1 root # waifor completion of thdeletion command sleep 1 # removthnetem kernel module /sbin/rmmod sch_netem sleep 1 # show qdisc configuration $TC -s qdisc show dev eth1 # show filter configuration $TC -s filter show dev eth1 ARCHIVES: --------- I may supply orequesthe experimental 2.6.9 and 2.6.11 deterministic pruning patches. I wish this simplpiecof code will help you. Yours sincerely, Mr. LaurenMARIE Frosimon amullis.co.uk Thu Sep 20 05:19:58 2007 From: simoamullis.co.uk (Simon Mullis) Date: Thu Sep 20 05:20:02 2007 Subject: newbiquestion: Adding fixed delay to thredifferent flows based oDSCP / ToS field Message-ID: <23d7e1bb0709200519s5df454d2l37d6d3331b495f45@xxxxxxxxxxxxxx> Hi All, I'new to tc / neteso apologies if any of what follows is daft... I would likto add latency to thredifferent incoming flows based on their DSCP markings for somtesting I'doing. tc qdisc del dev eth2 root tc qdisc add dev eth2 roohandl1: prio bands 10 tc qdisc add dev eth2 paren1:1 handl10: netem delay 20ms tc filter add dev eth2 protocol ip paren1:0 prio 1 u32 match ip tos 0x10 flowid 10:1 tc qdisc add dev eth2 paren1:2 handl20: netem delay 50ms tc filter add dev eth2 protocol ip paren1:0 prio 2 u32 match ip tos 0x18 flowid 20:2 tc qdisc add dev eth2 paren1:3 handl30: netem delay 100ms tc filter add dev eth2 protocol ip paren1:0 prio 2 u32 match ip tos 0x26 flowid 30:3 tc qdisc add dev eth2 paren1:4 handl40: pfifo tc filter add dev eth2 protocol ip paren1:0 prio 2 u32 match ip tos 0x00 flowid 40:4 Is this thrighway to nest? So, anything without any DSCP marking aall will havno added delay... Thanks iadvance SM -- SimoMullis _________________ simon@xxxxxxxxxxxx Froshemminger alinux-foundation.org Thu Sep 20 11:29:02 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: Thu Sep 20 11:29:10 2007 Subject: newbiquestion: Adding fixed delay to thredifferent flows based oDSCP / ToS field In-Reply-To: <23d7e1bb0709200519s5df454d2l37d6d3331b495f45@xxxxxxxxxxxxxx> References: <23d7e1bb0709200519s5df454d2l37d6d3331b495f45@xxxxxxxxxxxxxx> Message-ID: <20070920112902.2439a4ff@freepuppy.rosehill> OThu, 20 Sep 2007 14:19:58 +0200 "SimoMullis" <simon@xxxxxxxxxxxx> wrote: > Hi All, > > I'new to tc / neteso apologies if any of what follows is daft... > > I would likto add latency to thredifferent incoming flows based on > their DSCP markings for somtesting I'doing. > > tc qdisc del dev eth2 root > tc qdisc add dev eth2 roohandl1: prio bands 10 > tc qdisc add dev eth2 paren1:1 handl10: netem delay 20ms > tc filter add dev eth2 protocol ip paren1:0 prio 1 u32 match ip tos > 0x10 flowid 10:1 > tc qdisc add dev eth2 paren1:2 handl20: netem delay 50ms > tc filter add dev eth2 protocol ip paren1:0 prio 2 u32 match ip tos > 0x18 flowid 20:2 > tc qdisc add dev eth2 paren1:3 handl30: netem delay 100ms > tc filter add dev eth2 protocol ip paren1:0 prio 2 u32 match ip tos > 0x26 flowid 30:3 > tc qdisc add dev eth2 paren1:4 handl40: pfifo > tc filter add dev eth2 protocol ip paren1:0 prio 2 u32 match ip tos > 0x00 flowid 40:4 > > Is this thrighway to nest? So, anything without any DSCP marking > aall will havno added delay... > > Thanks iadvance > > SM > Should work. Try testing it, with differenTOS values. You should sedifferenping times, and the stats will go up with: tc -s qdisc ls You don'need an explicipfifo for non-netem queue's. That is what priority does anyway. -- StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx>
