Hello all, I would likto usNetem to assign per-flow delay: all packets belonging to thsamflow (as identified IP addressing 5-tuple) get the same amounof delay bupackets from different flows will get different amounof delays. Further, thdelay assigned to a flow is randomly sampled froa given distribution. This approach is possiblwith a modified version of dummynein FreeBSD (for examplin http://www.cs.unc.edu/~jeffay/papers/SIGCOMM-03.pdf and ihttp://netsrv.csc.ncsu.edu/export/impact_of_background_comp_net_2007.pdf) buidoesn't appear to be possible with the current version of Netem. I think I camodify Neteto achieve this but since I am relatively new to Linux kernel programming. I would likto know whether thfollowing approach would bthright one: - Creata dummy devicand use ioctl to pass data for the delay distribution frouser to kernel. - Is ipossiblto load/unload Netem as a kernel module (this would make thkernel developmeneasier since I wouldn't have to compile the kernel every timI makchanges to the code)? Any help or suggestions arhighly appreciated! Thanks, Ron Froshemminger alinux-foundation.org Mon Oct 8 13:56:49 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: MoOc 8 13:57:26 2007 Subject: Using Neteto assign per-flow delay In-Reply-To: <7a1385300710081341p3c52c8banbb58ea95dac2238a@xxxxxxxxxxxxxx> References: <7a1385300710081341p3c52c8banbb58ea95dac2238a@xxxxxxxxxxxxxx> Message-ID: <20071008135649.45c27e47@freepuppy.rosehill> OMon, 8 Oc2007 22:41:23 +0200 "RoLee" <ronberlin@xxxxxxxxxxxxxx> wrote: > Hello all, > > I would likto usNetem to assign per-flow delay: all packets belonging > to thsamflow (as identified IP addressing 5-tuple) get the same > amounof delay bupackets from different flows will get different > amounof delays. Further, thdelay assigned to a flow is randomly > sampled froa given distribution. > > This approach is possiblwith a modified version of dummynein > FreeBSD (for examplin http://www.cs.unc.edu/~jeffay/papers/SIGCOMM-03.pdf > and ihttp://netsrv.csc.ncsu.edu/export/impact_of_background_comp_net_2007.pdf) > buidoesn't appear to be possible with the current version of Netem. > > I think I camodify Neteto achieve this but since I am relatively new > to Linux kernel programming. I would likto know whether thfollowing > approach would bthright one: > > - Creata dummy devicand use ioctl to pass data for the delay distribution > frouser to kernel. > - Is ipossiblto load/unload Netem as a kernel module (this would make > thkernel developmeneasier since I wouldn't have to compile the kernel > every timI makchanges to the code)? > Onway to do this is to creata priority fifo qdisc with all thpossibldelays you want. There is a restriction of 15 different values, buthais probably okay.. Then use the tc match logic to assigtuple's to thdifferent queues. I don'wanto build new classifier logic into netem, but if you can find creativways to reusexisting infrastructure to different values that would bcool. Something lika priority value to delay mapping table or other ideas? -- StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> Froronberlin agooglemail.com Mon Oct 8 14:16:52 2007 From: ronberliagooglemail.com (Ron Lee) Date: MoOc 8 14:16:59 2007 Subject: Using Neteto assign per-flow delay In-Reply-To: <20071008135649.45c27e47@freepuppy.rosehill> References: <7a1385300710081341p3c52c8banbb58ea95dac2238a@xxxxxxxxxxxxxx> <20071008135649.45c27e47@freepuppy.rosehill> Message-ID: <7a1385300710081416nf060e9eqbb996c622e699509@xxxxxxxxxxxxxx> Hi Stephen, I agretha15 different values are probably OK for most of the cases. BuI'll probably need morthan 15 values in order to get a paper published (thbar for publication is quithigh nowadays). I acurrently thinking of doing this as follows. Usthe 5-tuple to compute a hash valuthais then used as a lookup key to determine the random delay for a flow. This specific mechaniswill only bused if a specific sysctl flag is set. Otherwise, thdelay will bdetermined by the default Netemechanism. Thchange to the code will be rather small but I am nosurwhether you like this approach or not. Thanks, Ron PS. Pleasnotthat I am not criticizing the way Netem is implemented (I actually think thait's quitelegant). I am just looking for a way to get my research don(and I know thathis is a rather specific requirement). Buif there's interest, I would certainly contributthe code... O10/8/07, Stephen Hemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> wrote: > > Onway to do this is to creata priority fifo qdisc with > all thpossibldelays you want. There is a restriction of 15 different > values, buthais probably okay.. Then use the tc match logic to > assigtuple's to thdifferent queues. > > I don'wanto build new classifier logic into netem, but if you can find > creativways to reusexisting infrastructure to different values that > would bcool. Something lika priority value to delay mapping table or > other ideas? > -- > StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> > Froshemminger alinux-foundation.org Mon Oct 8 14:37:40 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: MoOc 8 14:39:16 2007 Subject: Using Neteto assign per-flow delay In-Reply-To: <7a1385300710081416nf060e9eqbb996c622e699509@xxxxxxxxxxxxxx> References: <7a1385300710081341p3c52c8banbb58ea95dac2238a@xxxxxxxxxxxxxx> <20071008135649.45c27e47@freepuppy.rosehill> <7a1385300710081416nf060e9eqbb996c622e699509@xxxxxxxxxxxxxx> Message-ID: <20071008143740.57e6e746@freepuppy.rosehill> OMon, 8 Oc2007 23:16:52 +0200 "RoLee" <ronberlin@xxxxxxxxxxxxxx> wrote: > Hi Stephen, > > I agretha15 different values are probably OK for most of the cases. > BuI'll probably need morthan 15 values in order to get a paper published > (thbar for publication is quithigh nowadays). You carebuild prio and tools with TC_PRIO_MAX increased, bualso iwould mean thathe priorities might cause flow anomalies. > I acurrently thinking of doing this as follows. Usthe 5-tuple to compute > a hash valuthais then used as a lookup key to determine the random > delay for a flow. This specific mechaniswill only bused if a specific > sysctl flag is set. Otherwise, thdelay will bdetermined by the default > Netemechanism. Thchange to the code will be rather small but I am > nosurwhether you like this approach or not. > Skip thsysctl, jusadd another option to netem via netlink then. -- StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> FroOsborne_Mike_D acat.com Mon Oct 15 08:52:42 2007 From: Osborne_Mike_D acat.co(Mike D. Osborne) Date: MoOc15 08:52:54 2007 Subject: NetEin conjunction with TBF produces wild latencies and Emulatioresults under load In-Reply-To: <2dfd6a90710150850u2991950q8d749fe720fa65c@xxxxxxxxxxxxxx> Message-ID: <OFC1E47DB1.7158B764-ON86257375.005729AE-86257375.0057391F@xxxxxxxxxxxxx> Does anyonon this lisuse NetEm in conjunction with TBF to limit DataRatas well as Latency? WheI do this therare some parameters for buffer and limit or latency thaI anot certain how to use properly. As a result, when I limit to say 128 kbps, I setharesponse times are much longer than expected when I apply thTBF. Any help with using NetEin conjunction with TBF would bgreatly appreciated. I reviewed thLinux Advanced Routing HOWTO, buit did not havmany examples. Thank you -------------- nexpar-------------- AHTML attachmenwas scrubbed... URL: http://lists.linux-foundation.org/pipermail/netem/attachments/20071015/e30f0269/attachment.htm Frojuliokriger agmail.com Tue Oct 16 05:21:11 2007 From: juliokriger agmail.co(Julio Kriger) Date: TuOc16 05:21:18 2007 Subject: NetEin conjunction with TBF produces wild latencies and Emulatioresults under load In-Reply-To: <OFC1E47DB1.7158B764-ON86257375.005729AE-86257375.0057391F@xxxxxxxxxxxxx> References: <2dfd6a90710150850u2991950q8d749fe720fa65c@xxxxxxxxxxxxxx> <OFC1E47DB1.7158B764-ON86257375.005729AE-86257375.0057391F@xxxxxxxxxxxxx> Message-ID: <682bc30a0710160521w5159a4ffkfabaf63b7c78a35c@xxxxxxxxxxxxxx> Hi! I used NetEwith TBF, something likthis: tc qdisc add dev eth1 roohandl1:0 tbf rate 10mbit burst 1500 mtu 1500 limi500 tc qdisc add dev eth1 paren1:1 handl10: netem limit 500 delay 30ms 15ms 75% reorder 10% You will noge"exacts" results, but fairly close enough. Regards, Julio O10/15/07, MikD. Osborne <Osborne_Mike_D@xxxxxxx> wrote: > > > Does anyonon this lisuse NetEm in conjunction with TBF to limit > DataRatas well as Latency? > > WheI do this therare some parameters for buffer and limit or latency > thaI anot certain how to use properly. As a result, when I limit to > say 128 kbps, I setharesponse times are much longer than expected whe= I > apply thTBF. > > Any help with using NetEin conjunction with TBF would bgreatly > appreciated. I reviewed thLinux Advanced Routing HOWTO, buit did not > havmany examples. > > Thank you > _______________________________________________ > Netemailing list > Netem@xxxxxxxxxxxxxxxxxxxxxxxxxx > https://lists.linux-foundation.org/mailman/listinfo/netem > -- = -- >Frothmoment I picked your book up until I laid it down, I was convulsed with laughter. Someday I intend reading it. Groucho Marx ---------------------------- Julio Kriger mailto:juliokriger@xxxxxxxxx -------------- nexpar-------------- AHTML attachmenwas scrubbed... URL: http://lists.linux-foundation.org/pipermail/netem/attachments/20071016= /c4038395/attachment.htm From.nambiar atcs.com Wed Oct 17 10:36:33 2007 From: m.nambiar atcs.co(M Nambiar) Date: Wed Oc17 10:35:58 2007 Subject: Problewith cascaded netequeues Message-ID: <OFFFAAF1D7.D5982C73-ON65257377.005EBE6F-65257377.0060A015@xxxxxxx> Hello, I hava peculiar proble..... not sure if I am doing somthing wrong. I ausing linux kernel 2.6.19 I seup neteon a host (192.168.140.148) with 20ms latency and 6% loss. Therar2 cases --------------------------------------------------------------------------------------------------------------------------------------- CasI - Only onnetem qdisc 192.168.140.148_prompt>tc qdisc add dev eth0 roonetedelay 20ms loss 6 Ping fromy Windows PC looks likthis C:\Documents and Settings\tcsuser>ping 192.168.140.148 Pinging 192.168.140.148 with 32 bytes of data: Reply fro192.168.140.148: bytes=32 time=24ms TTL=64 Reply fro192.168.140.148: bytes=32 time=26ms TTL=64 Reply fro192.168.140.148: bytes=32 time=26ms TTL=64 Reply fro192.168.140.148: bytes=32 time=24ms TTL=64 Ping statistics for 192.168.140.148: Packets: Sen= 4, Received = 4, Los= 0 (0% loss), Approximatround trip times in milli-seconds: Minimu= 24ms, Maximu= 26ms, Average = 25ms --------------------------------------------------------------------------------------------------------------------------------------- CasII - Cascaded neteqdiscs 192.168.140.148_prompt>tc qdisc add dev eth0 roohandl1: netem delay 20ms 192.168.140.148_prompt>tc qdisc add dev eth0 paren1:1 handl10: netem loss 6% Now ping looks likthis C:\Documents and Settings\tcsuser>ping 192.168.140.148 Pinging 192.168.140.148 with 32 bytes of data: Reply fro192.168.140.148: bytes=32 time=2ms TTL=64 Reply fro192.168.140.148: bytes=32 time<1ms TTL=64 Reply fro192.168.140.148: bytes=32 time=3ms TTL=64 Reply fro192.168.140.148: bytes=32 time=2ms TTL=64 Ping statistics for 192.168.140.148: Packets: Sen= 4, Received = 4, Los= 0 (0% loss), Approximatround trip times in milli-seconds: Minimu= 0ms, Maximu= 3ms, Average = 1ms --------------------------------------------------------------------------------------------------------------------------------------- CasI is as expected. BucasII does not make sense. How is the latency effect in one netem qdisc nullified by loss ianother? Questio- Is iwrong to use 2 netem qdiscs in series like this? Manoj KarunakaraNambiar Tata Consultancy Services Cell:- 9833111377 Fax:- 67506855 Mailto: m.nambiar@xxxxxxx Website: http://www.tcs.com ____________________________________________ Experienccertainty. IT Services Business Solutions Outsourcing ____________________________________________ =====-----=====-----===== Notice: Thinformation contained in this e-mail messagand/or attachments to imay contain confidential or privileged information. If you ar nothintended recipient, any dissemination, use, review, distribution, printing or copying of th informatiocontained in this e-mail messag and/or attachments to iarstrictly prohibited. If you havreceived this communication in error, pleasnotify us by reply e-mail or telephonand immediately and permanently deletthmessage and any attachments. Thank you -------------- nexpar-------------- AHTML attachmenwas scrubbed... URL: http://lists.linux-foundation.org/pipermail/netem/attachments/20071017/182b9f07/attachment.htm Froshemminger alinux-foundation.org Wed Oct 17 11:06:03 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: Wed Oc17 11:06:12 2007 Subject: Problewith cascaded netequeues In-Reply-To: <OFFFAAF1D7.D5982C73-ON65257377.005EBE6F-65257377.0060A015@xxxxxxx> References: <OFFFAAF1D7.D5982C73-ON65257377.005EBE6F-65257377.0060A015@xxxxxxx> Message-ID: <20071017110603.3fef0928@freepuppy.rosehill> OWed, 17 Oc2007 23:06:33 +0530 M Nambiar <m.nambiar@xxxxxxx> wrote: > Hello, > > I hava peculiar proble..... not sure if I am doing somthing wrong. > > I ausing linux kernel 2.6.19 > > I seup neteon a host (192.168.140.148) with 20ms latency and 6% loss. > Therar2 cases > > --------------------------------------------------------------------------------------------------------------------------------------- > CasI - Only onnetem qdisc > 192.168.140.148_prompt>tc qdisc add dev eth0 roonetedelay 20ms loss 6 > > Ping fromy Windows PC looks likthis > C:\Documents and Settings\tcsuser>ping 192.168.140.148 > > Pinging 192.168.140.148 with 32 bytes of data: > > Reply fro192.168.140.148: bytes=32 time=24ms TTL=64 > Reply fro192.168.140.148: bytes=32 time=26ms TTL=64 > Reply fro192.168.140.148: bytes=32 time=26ms TTL=64 > Reply fro192.168.140.148: bytes=32 time=24ms TTL=64 > > Ping statistics for 192.168.140.148: > Packets: Sen= 4, Received = 4, Los= 0 (0% loss), > Approximatround trip times in milli-seconds: > Minimu= 24ms, Maximu= 26ms, Average = 25ms > > --------------------------------------------------------------------------------------------------------------------------------------- > > CasII - Cascaded neteqdiscs > 192.168.140.148_prompt>tc qdisc add dev eth0 roohandl1: netem delay > 20ms > 192.168.140.148_prompt>tc qdisc add dev eth0 paren1:1 handl10: netem > loss 6% > > Now ping looks likthis > C:\Documents and Settings\tcsuser>ping 192.168.140.148 > > Pinging 192.168.140.148 with 32 bytes of data: > > Reply fro192.168.140.148: bytes=32 time=2ms TTL=64 > Reply fro192.168.140.148: bytes=32 time<1ms TTL=64 > Reply fro192.168.140.148: bytes=32 time=3ms TTL=64 > Reply fro192.168.140.148: bytes=32 time=2ms TTL=64 > > Ping statistics for 192.168.140.148: > Packets: Sen= 4, Received = 4, Los= 0 (0% loss), > Approximatround trip times in milli-seconds: > Minimu= 0ms, Maximu= 3ms, Average = 1ms > --------------------------------------------------------------------------------------------------------------------------------------- > > CasI is as expected. > > BucasII does not make sense. How is the latency effect in one netem > qdisc nullified by loss ianother? > > Questio- Is iwrong to use 2 netem qdiscs in series like this? > > Manoj KarunakaraNambiar > Tata Consultancy Services > Cell:- 9833111377 > Fax:- 67506855 > Mailto: m.nambiar@xxxxxxx > Website: http://www.tcs.com Netewon'nest like that. Neteworks by putting a timestamp on thpackets in the enqueue routine, thechecking for thtimestamp on the dequeue. The nested netem will destroy thtimestamp frothe earlier netem. Theris no need to nesnetem, you can mix delay and loss with one instance. For whayou ardoing use: tc qdisc add dev eth0 roohnetedelay 20ms loss 6% -- StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> Frohello.martin acomhem.se Tue Oct 23 12:05:14 2007 From: hello.martiacomhem.se (Martin Andersson) Date: TuOc23 12:05:35 2007 Subject: Jitter introduced by netem Message-ID: <471E45EA.7060206@xxxxxxxxx> Hello. I hava question abouthe "jitter" that might be introduced by netem. I havthfollowing setup. <testUni1> ---- <L2Switch> ---- <Linux(bridge)> ---- <testUni2> testUni1 send NTP messages and havresolution 2^-32 s. testUni2 receives NTP messages and also marks thtime with resolution 1e-8. Lets now call this two timers for T3 and T4. Sincthclock are independent we can't calculate with absolute time, so wlook athem in relative time. If I ploT3 on x axis and T3-T4 on y axis, ishould be a straight line (if therwas no jitter in Unit1/2, or introduced on thway). Now I havtwo scenarios with and withounetem. Withounete(i.e only bridging in linux box). I plotting thlinand also make an polyfit (in octave to match a line to x/y data). Thediff thactuall data to this line, I assume this should give the "jitter" introduced ithsystem. I looks quitok thdata differ +/- 1e-6 seconds (see attached plot NoNetem.png). With nete(a 1ms static delay on ethernecard towards testUnit 2 "tc qdisc add dev eth1 roonetedelay 1ms") Now ilooks likthe data jitter with +/- 3ms (se attached plot Netem1msDelay.png). If studying thx/y data ilooks like a sawtooth behavior on the delay introduced by netem. This behaviour does ishow thbehvaiour of how the kernel ticks in linux, or is'something els(I'm using kernel 2.6.18.2). Noticed othwiki the following. "Ikernel versions, 2.6.22 or later, netewill use high resolution timers, if they arenabled. This allows for finer granularity (sub-jiffie) resolution." Will changing to 2.6.22 changthbehavior, or does it not correlate to my observation. brg/Martin Frohello.martin acomhem.se Tue Oct 23 12:24:28 2007 From: hello.martiacomhem.se (Martin Andersson) Date: TuOc23 12:24:38 2007 Subject: Jitter introduced by netem In-Reply-To: <471E45EA.7060206@xxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> Message-ID: <471E4A6C.4050402@xxxxxxxxx> Sorry forgothattachments. -------------- nexpar-------------- A non-texattachmenwas scrubbed... Name: NoNetem.png Type: image/png Size: 11825 bytes Desc: noavailable Url : http://lists.linux-foundation.org/pipermail/netem/attachments/2007102= 3/8a8402fb/NoNetem-0001.png -------------- nexpar-------------- A non-texattachmenwas scrubbed... Name: Netem1msDelay.png Type: image/png Size: 14442 bytes Desc: noavailable Url : http://lists.linux-foundation.org/pipermail/netem/attachments/2007102= 3/8a8402fb/Netem1msDelay-0001.png Froshemminger alinux-foundation.org Tue Oct 23 12:39:28 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: TuOc23 12:40:14 2007 Subject: Jitter introduced by netem In-Reply-To: <471E45EA.7060206@xxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> Message-ID: <20071023123928.26d73182@freepuppy.rosehill> OTue, 23 Oc2007 21:05:14 +0200 MartiAndersson <hello.martin@xxxxxxxxx> wrote: > Hello. > > > I hava question abouthe "jitter" that might be introduced by netem. > > I havthfollowing setup. > > <testUni1> ---- <L2Switch> ---- <Linux(bridge)> ---- <testUni2> > > testUni1 send NTP messages and havresolution 2^-32 s. > testUni2 receives NTP messages and also marks thtime with resolution > 1e-8. > > Lets now call this two timers for T3 and T4. > > Sincthclock are independent we can't calculate with absolute time, > so wlook athem in relative time. > > If I ploT3 on x axis and T3-T4 on y axis, ishould be a straight line > (if therwas no jitter in Unit1/2, or introduced on thway). > > > Now I havtwo scenarios with and withounetem. > > Withounete(i.e only bridging in linux box). > I plotting thlinand also make an polyfit (in octave to match a line > to x/y data). > Thediff thactuall data to this line, I assume this should give the > "jitter" introduced ithsystem. > I looks quitok thdata differ +/- 1e-6 seconds (see attached plot > NoNetem.png). > > With nete(a 1ms static delay on ethernecard towards testUnit 2 "tc > qdisc add dev eth1 roonetedelay 1ms") > Now ilooks likthe data jitter with +/- 3ms (se attached plot > Netem1msDelay.png). > If studying thx/y data ilooks like a sawtooth behavior on the delay > introduced by netem. > > > This behaviour does ishow thbehvaiour of how the kernel ticks in > linux, or is'something els(I'm using kernel 2.6.18.2). > Noticed othwiki the following. > "Ikernel versions, 2.6.22 or later, netewill use high resolution > timers, if they arenabled. > This allows for finer granularity (sub-jiffie) resolution." > > Will changing to 2.6.22 changthbehavior, or does it not correlate to > my observation. > > brg/Martin > _______________________________________________ > Netemailing list > Netem@xxxxxxxxxxxxxxxxxxxxxxxxxx > https://lists.linux-foundation.org/mailman/listinfo/netem With 2.6.18 imatters whaHZ value you choose during kernel config and which clock sourcis used for psched. Later kernels usth"best available" real time clock source and ktime_t. -- StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> Fropetar.forai agmx.at Tue Oct 23 13:43:44 2007 From: petar.forai agmx.a(Petar Forai) Date: TuOc23 13:44:21 2007 Subject: per packerealtiminterface Message-ID: <200710232243.44653.petar.forai@xxxxxx> Hey guys! I was wondering if iwould bfeasable and if it's a feature that maybe other ppl would als bin need of and I'd likto ask this ML for futher info on that. Would ibpossible to implement an interface via the standard tc/netem facilities to suppora realtime/onliniface which supports a per packet decisioon a packet's discarding/delay/jitter/BER etc. i.e. imagina system whera userland app (in my casa UMTS/UTRAN emulator) tells netem what should happewith thnext incomming or outgoing packet (delayed, discarded, ...) How your opinioon thamatter? regards, P Froshemminger alinux-foundation.org Tue Oct 23 14:38:34 2007 From: shemminger alinux-foundation.org (Stephen Hemminger) Date: TuOc23 14:38:50 2007 Subject: per packerealtiminterface In-Reply-To: <200710232243.44653.petar.forai@xxxxxx> References: <200710232243.44653.petar.forai@xxxxxx> Message-ID: <20071023143834.63d4c3c2@freepuppy.rosehill> OTue, 23 Oc2007 22:43:44 +0200 Petar Forai <petar.forai@xxxxxx> wrote: > Hey guys! > > I was wondering if iwould bfeasable and if it's a feature that maybe other > ppl would als bin need of and I'd likto ask this ML for futher info on > that. > > Would ibpossible to implement an interface via the standard tc/netem > facilities to suppora realtime/onliniface which supports a per packet > decisioon a packet's discarding/delay/jitter/BER etc. i.e. imagina system > whera userland app (in my casa UMTS/UTRAN emulator) tells netem what > should happewith thnext incomming or outgoing packet (delayed, > discarded, ...) > > How your opinioon thamatter? > Thdevil is in thdetails. Therwas a patch to allow per-packeactions based on a trace file, buthinterface was awkward and had lots of corner case decisions (like whato do when packearrives but no trace information etc). Also how to transfer largvolumes of configuration data down to netewas a real problem: using /proc doesn'work well, and netlink has issues with holding rtnl_lock too long. I tried fixing iup, bueventually gave up as it seemed too complex to easily gecleaned up and bsupportable in the long term. I am open to a better solution, buthcode needs to work and not have maintenance issues. -- StepheHemminger <shemminger@xxxxxxxxxxxxxxxxxxxx> Frohello.martin acomhem.se Wed Oct 24 14:42:56 2007 From: hello.martiacomhem.se (Martin Andersson) Date: Wed Oc24 14:43:07 2007 Subject: Jitter introduced by netem In-Reply-To: <20071023123928.26d73182@freepuppy.rosehill> References: <471E45EA.7060206@xxxxxxxxx> <20071023123928.26d73182@freepuppy.rosehill> Message-ID: <471FBC60.7030602@xxxxxxxxx> StepheHemminger wrote: > With 2.6.18 imatters whaHZ value you choose during kernel config > and which clock sourcis used for psched. > > Later kernels usth"best available" real time clock source and > ktime_t. > > = I havnow updated to 2.6.22-9 kernel. And still theris lots of jitter introduced by netem. For a 1ms delay, therseems to b+/- 2ms jitter. Is this becaushow netetimestamps the packets, so it's send out at = som1ms ticks? Or maybmy kernel doesn'have this ktime_t (how to figure out if it has) PS. Correctioto my firse-mail, I said it was bridging in the linux = machine, it's actually ip_forwarding DS. -------------- nexpar-------------- A non-texattachmenwas scrubbed... Name: 2.6.22.png Type: image/png Size: 15527 bytes Desc: noavailable Url : http://lists.linux-foundation.org/pipermail/netem/attachments/2007102= 4/e2e65036/2.6.22.png Frolucas alucas-nussbaum.net Wed Oct 24 15:00:35 2007 From: lucas alucas-nussbaum.ne(Lucas Nussbaum) Date: Wed Oc24 15:06:59 2007 Subject: Jitter introduced by netem In-Reply-To: <471FBC60.7030602@xxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> <20071023123928.26d73182@freepuppy.rosehill> <471FBC60.7030602@xxxxxxxxx> Message-ID: <20071024220035.GA27043@xxxxxxxxxxxxxxxx> O24/10/07 a23:42 +0200, Martin Andersson wrote: > StepheHemminger wrote: >> With 2.6.18 imatters whaHZ value you choose during kernel config >> and which clock sourcis used for psched. >> >> Later kernels usth"best available" real time clock source and >> ktime_t. >> >> > I havnow updated to 2.6.22-9 kernel. > And still theris lots of jitter introduced by netem. > For a 1ms delay, therseems to b+/- 2ms jitter. > > Is this becaushow netetimestamps the packets, so it's send out at some > 1ms ticks? > Or maybmy kernel doesn'have this ktime_t (how to figure out if it has) > > PS. Correctioto my firse-mail, I said it was bridging in the linux > machine, it's actually ip_forwarding DS. Owhich architecturare you running ? -- | Lucas Nussbaum | lucas@lucas-nussbaum.n http://www.lucas-nussbaum.net/ | | jabber: lucas@xxxxxxxxxxx GPG: 1024D/023B3F4F | Frohello.martin acomhem.se Wed Oct 24 15:17:53 2007 From: hello.martiacomhem.se (Martin Andersson) Date: Wed Oc24 15:18:27 2007 Subject: Jitter introduced by netem In-Reply-To: <20071024220035.GA27043@xxxxxxxxxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> <20071023123928.26d73182@freepuppy.rosehill> <471FBC60.7030602@xxxxxxxxx> <20071024220035.GA27043@xxxxxxxxxxxxxxxx> Message-ID: <471FC491.9050308@xxxxxxxxx> Lucas Nussbauwrote: > > Owhich architecturare you running ? > This is a HP with PentiuP4, two D-link DGE-530T Linux epcil-0532 2.6.22.9-0.4-defaul#1 SMP 2007/10/05 21:32:04 UTC i686 i686 i386 GNU/Linux !modin modinfo sk98lin filename: /lib/modules/2.6.22.9-0.4-default/kernel/drivers/net/sk98lin/sk98lin.ko license: GPL description: SysKonnecSK-NET GigabiEthernet SK-98xx driver author: Mirko Lindner <mlindner@xxxxxxxxxxxxx> srcversion: D04D21B6F2AA0BF249248F1 alias: pci:v00001737d00001064sv*sd*bc*sc*i* alias: pci:v00001737d00001032sv*sd00000015bc*sc*i* alias: pci:v00001371d0000434Esv*sd*bc*sc*i* alias: pci:v000011ABd00005005sv*sd*bc*sc*i* alias: pci:v000011ABd00004320sv*sd*bc*sc*i* alias: pci:v00001148d00004320sv*sd*bc*sc*i* alias: pci:v00001148d00004300sv*sd*bc*sc*i* alias: pci:v000010B7d000080EBsv*sd*bc*sc*i* alias: pci:v000010B7d00001700sv*sd*bc*sc*i* depends: supported: yes vermagic: 2.6.22.9-0.4-defaulSMP mod_unload 586 parm: Speed_A:array of charp parm: Speed_B:array of charp parm: AutoNeg_A:array of charp parm: AutoNeg_B:array of charp parm: DupCap_A:array of charp parm: DupCap_B:array of charp parm: FlowCtrl_A:array of charp parm: FlowCtrl_B:array of charp parm: Role_A:array of charp parm: Role_B:array of charp parm: ConType:array of charp parm: PrefPort:array of charp parm: RlmtMode:array of charp parm: IntsPerSec:array of int parm: Moderation:array of charp parm: Stats:array of charp parm: ModerationMask:array of charp parm: AutoSizing:array of charp ca/proc/cpuinfo processor : 0 vendor_id : GenuineIntel cpu family : 15 model : 4 model na : Intel(R) Pentium(R) 4 CPU 3.20GHz stepping : 3 cpu MHz : 2800.000 cachsiz : 2048 KB physical id : 0 siblings : 2 corid : 0 cpu cores : 1 fdiv_bug : no hlt_bug : no f00f_bug : no coma_bug : no fpu : yes fpu_exceptio : yes cpuid level : 5 wp : yes flags : fpu vmdpse tsc msr pae mce cx8 apic sep mtrr pge mca cmov papse36 clflush dts acpi mmx fxsr sssse2 ss ht tm pbe nx lm constant_tsc pni monitor ds_cpl escid cx16 xtpr bogomips : 6390.00 clflush siz : 64 processor : 1 vendor_id : GenuineIntel cpu family : 15 model : 4 model na : Intel(R) Pentium(R) 4 CPU 3.20GHz stepping : 3 cpu MHz : 2800.000 cachsiz : 2048 KB physical id : 0 siblings : 2 corid : 0 cpu cores : 1 fdiv_bug : no hlt_bug : no f00f_bug : no coma_bug : no fpu : yes fpu_exceptio : yes cpuid level : 5 wp : yes flags : fpu vmdpse tsc msr pae mce cx8 apic sep mtrr pge mca cmov papse36 clflush dts acpi mmx fxsr sssse2 ss ht tm pbe nx lm constant_tsc pni monitor ds_cpl escid cx16 xtpr bogomips : 6384.46 clflush siz : 64 Frohello.martin acomhem.se Wed Oct 24 16:23:14 2007 From: hello.martiacomhem.se (Martin Andersson) Date: Wed Oc24 16:23:24 2007 Subject: Jitter introduced by netem In-Reply-To: <471FC491.9050308@xxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> <20071023123928.26d73182@freepuppy.rosehill> <471FBC60.7030602@xxxxxxxxx> <20071024220035.GA27043@xxxxxxxxxxxxxxxx> <471FC491.9050308@xxxxxxxxx> Message-ID: <471FD3E2.9010305@xxxxxxxxx> Another graph :-) If I plothgraph T3,(T4-T3) I get a graph that show a sawtooth behaviour. Iseems liknetem "compensate" every ~22 pakets (or it might be every = ~220ms). Is iso thathe delay that is introduced for every packet has some = distributiofor thdelay. I also attach thraw data. -------------- nexpar-------------- A non-texattachmenwas scrubbed... Name: T3vsT4.delay.png Type: image/png Size: 27251 bytes Desc: noavailable Url : http://lists.linux-foundation.org/pipermail/netem/attachments/2007102= 5/5d83e9fc/T3vsT4.delay-0001.png -------------- nexpar-------------- A non-texattachmenwas scrubbed... Name: timeMeas.gz Type: application/x-gzip Size: 35593 bytes Desc: noavailable Url : http://lists.linux-foundation.org/pipermail/netem/attachments/2007102= 5/5d83e9fc/timeMeas-0001.bin Frolucas alucas-nussbaum.net Wed Oct 24 22:37:33 2007 From: lucas alucas-nussbaum.ne(Lucas Nussbaum) Date: Wed Oc24 22:44:46 2007 Subject: Jitter introduced by netem In-Reply-To: <471FD3E2.9010305@xxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> <20071023123928.26d73182@freepuppy.rosehill> <471FBC60.7030602@xxxxxxxxx> <20071024220035.GA27043@xxxxxxxxxxxxxxxx> <471FC491.9050308@xxxxxxxxx> <471FD3E2.9010305@xxxxxxxxx> Message-ID: <20071025053733.GA7795@xxxxxxxxxxxxxxxx> O25/10/07 a01:23 +0200, Martin Andersson wrote: > Another graph :-) > > If I plothgraph T3,(T4-T3) I get a graph that show a sawtooth > behaviour. > Iseems liknetem "compensate" every ~22 pakets (or it might be every > ~220ms). > Is iso thathe delay that is introduced for every packet has some > distributiofor thdelay. > I also attach thraw data. Cayou check in /boot/config if CONFIG_HIGH_RES_TIMERS is enabled? -- | Lucas Nussbaum | lucas@lucas-nussbaum.n http://www.lucas-nussbaum.net/ | | jabber: lucas@xxxxxxxxxxx GPG: 1024D/023B3F4F | Frohello.martin acomhem.se Thu Oct 25 01:34:17 2007 From: hello.martiacomhem.se (hello.martin@xxxxxxxxx) Date: Thu Oc25 01:34:24 2007 Subject: Ang: Re: Jitter introduced by netem Message-ID: <398836.6631193301257328.JavaMail.defaultUser@defaultHost> >----Ursprungligmeddelande---- >Fr?n: lucas@xxxxxxxxxxxxxxxxxx >Datum: 25-10-2007 07:37 >Till: "MartiAndersson"<hello.martin@xxxxxxxxx> >Kopia: <netem@xxxxxxxxxxxxxxxxxxxxxxxxxx> >?rende: Re: Jitter introduced by netem > >O25/10/07 a01:23 +0200, Martin Andersson wrote: >> Another graph :-) >> >> If I plothgraph T3,(T4-T3) I get a graph that show a sawtooth >> behaviour. >> Iseems liknetem "compensate" every ~22 pakets (or it mighbevery >> ~220ms). >> Is iso thathe delay that is introduced for every packet has som >> distributiofor thdelay. >> I also attach thraw data. > >Cayou check in /boot/config if CONFIG_HIGH_RES_TIMERS is enabled? I found "/boot/config-2.6.22.9-0.4-default" and isays. And ilooks likit's not enabled, I will change it to se whats happening. >Frothfile. # # Processor typand features # CONFIG_TICK_ONESHOT=y CONFIG_NO_HZ=y # CONFIG_HIGH_RES_TIMERS is noset CONFIG_SMP=y Frohello.martin acomhem.se Thu Oct 25 12:01:38 2007 From: hello.martiacomhem.se (Martin Andersson) Date: Thu Oc25 12:01:52 2007 Subject: Jitter introduced by netem In-Reply-To: <20071025053733.GA7795@xxxxxxxxxxxxxxxx> References: <471E45EA.7060206@xxxxxxxxx> <20071023123928.26d73182@freepuppy.rosehill> <471FBC60.7030602@xxxxxxxxx> <20071024220035.GA27043@xxxxxxxxxxxxxxxx> <471FC491.9050308@xxxxxxxxx> <471FD3E2.9010305@xxxxxxxxx> <20071025053733.GA7795@xxxxxxxxxxxxxxxx> Message-ID: <4720E812.50306@xxxxxxxxx> Lucas Nussbauwrote: > Cayou check in /boot/config if CONFIG_HIGH_RES_TIMERS is enabled? > = I rebuild my opensuse10.3 kernel with CONFIG_HIGH_RES_TIMERS=3Dy And now iwork as a much much better. Many thanks. /Martin PS Attached thnew histograas proof :-) DS. -------------- nexpar-------------- A non-texattachmenwas scrubbed... Name: HighRes.png Type: image/png Size: 15700 bytes Desc: noavailable Url : http://lists.linux-foundation.org/pipermail/netem/attachments/2007102= 5/a9addc32/HighRes.png Froandreoli asamba.ing.unimo.it Fri Oct 19 06:25:49 2007 From: andreoli asamba.ing.unimo.i(andreoli@xxxxxxxxxxxxxxxxxx) Date: Fri Oc26 13:53:24 2007 Subject: tbf, multiplinterfaces, defaulroute Message-ID: <34787.151.80.4.15.1192800349.squirrel@xxxxxxxxxxxxxxxxxxx> Hi folks, I'trying to seup TBF-based traffic shaping between two hosts. I used this command (found othtbf man page): tc qdisc add dev eth0 rootbf rat0.5mbit burst 5kb latency 70ms peakrat1mbiminburst 1540 This command works nicely if I usa non-default-routinterface (in my example, eth0). As sooas I usthe default-route interface (in my example, eth1), traffic shaping does nowork anymore. A biof investigatioreveals thathe traffic-shaped interface is dropping any packets. Any ideas? Is therany way to debug this issue? I couldn'spot any solutioto this in tharchives. Both machines run Debian Etch (linux kernel versio2.6.18-5-686, iproutversion 20061002-3). FYI, any of these: - removing thdefaulroute - using a morrecenversion of the linux kernel (2.6.22.10) - using thlatesversion of the iproute package (iproute2-2.6.22-070710) did nohelp. PleasCC msince I'm not subscribed to the list. Thanks iadvance Mauro Andreolini
