[Int-area] Stateless IPv4-in-IPv6 experiments
Dear all, An I-D about stateless IPv4-in-IPv6 experiments has been submitted. A URL for this Internet-Draft is: http://tools.ietf.org/html/draft-deng-aplusp-experiment-results-00 http://tools.ietf.org/html/draft-deng-aplusp-experiment-results-00 . A website introducing the detail of experiment results is also accessible via: http://opensourceaplusp.weebly.com/ http://opensourceaplusp.weebly.com/ BR, Xiaohong ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] draft-ietf-intarea-ipv4-id-update-02.txt
On Tue, Apr 5, 2011 at 16:44, Iljitsch van Beijnum iljit...@muada.comwrote: On 5 apr 2011, at 18:51, Joe Touch wrote: This is the idea that an ID value can't be repeated for 120 (or 240) seconds for the same 3-tuple. But why 120 seconds? Is there any reasonable way that a packet gets fragmented, and then one of the fragments is delayed so long than 65535 other packets are received in the interim? I can't think of anything. There are plenty of systems with path asymmetries with delays of tens of seconds. E.g., see Jim Getty's presentation this past IETF in TSVAREA. Can't find that real quick. You would enjoy a visit to http://www.bufferbloat.net Scott ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] draft-ietf-intarea-ipv4-id-update-02.txt
On 6 apr 2011, at 17:33, Scott Brim wrote: Jim Getty's presentation this past IETF in TSVAREA. Can't find that real quick. You would enjoy a visit to http://www.bufferbloat.net Ah, it starts at 25 minutes in on the recording for wednesday morning on channel 4, with the slides here: http://mirrors.bufferbloat.net/Talks/PragueIETF/IETFBloat7.pdf BTW, I'm not seeing the disasters he's describing, I only get 6 or 12 ms extra delay doing 100 Mbps transfers locally. But... none of this has ANYTHING to do with the issue at hand, unless someone can show me how packet 50 sits in the buffer for 100 or 1000 or what have you milliseconds and then packet 51 magically flies through the queue and arrives much earlier than packet 50. ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] draft-ietf-intarea-ipv4-id-update-02.txt
On Tue, 5 Apr 2011, Iljitsch van Beijnum wrote: - even if it happens, the timing differences between reordered packets are counted in microseconds, not even whole milliseconds I have to disagree with this. I always discourage people from using per-packet loadblanacing, but if done on a 2 meg connection (as an example), just clocking out a 1500 byte packet takes 6 milliseconds, so two packets can easily arrive several milliseconds in wrong order in this use-case. In addition to that, I don't see multiple second delays AND hundreds or thousands of packets per second happening at the same time. Slow networks could potentially do the former, fast networks the latter, but I can't think of anything that can do both. I have to agree with that, though. -- Mikael Abrahamssonemail: swm...@swm.pp.se ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] draft-ietf-intarea-ipv4-id-update-02.txt
On 6 apr 2011, at 18:57, Mikael Abrahamsson wrote: - even if it happens, the timing differences between reordered packets are counted in microseconds, not even whole milliseconds I have to disagree with this. I always discourage people from using per-packet loadblanacing, but if done on a 2 meg connection (as an example), just clocking out a 1500 byte packet takes 6 milliseconds, so two packets can easily arrive several milliseconds in wrong order in this use-case. I've been working on 100 Mbps links, obviously this is different if you run much slower, which can happen at the edge of the network. But not really in the core, these days, I would say. ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] draft-ietf-intarea-ipv4-id-update-02.txt
On 4/5/2011 1:44 PM, Iljitsch van Beijnum wrote: On 5 apr 2011, at 18:51, Joe Touch wrote: This is the idea that an ID value can't be repeated for 120 (or 240) seconds for the same 3-tuple. But why 120 seconds? Is there any reasonable way that a packet gets fragmented, and then one of the fragments is delayed so long than 65535 other packets are received in the interim? I can't think of anything. There are plenty of systems with path asymmetries with delays of tens of seconds. E.g., see Jim Getty's presentation this past IETF in TSVAREA. Can't find that real quick. But I'm doing some measurements currently, and what I'm seeing doesn't support that at all, for two reasons: The Internet is a very big place, and it's intended to work in most places, rather than just in the places any individual currently sees. - reordering (= the packets must be sent down different paths) is very rare for packets that belong to the same session, think single digit percentages As noted in my other post, that could be increasing with multiple disparate connections becoming the norm. - even if it happens, the timing differences between reordered packets are counted in microseconds, not even whole milliseconds You should measure 3G, and include a satellite path too. Both have longer latencies even without buffer bloat. In addition to that, I don't see multiple second delays AND hundreds or thousands of packets per second happening at the same time. Slow networks could potentially do the former, fast networks the latter, but I can't think of anything that can do both. That's common over satellite paths, and also common in the buffer bloat case. As is the note that upper layers should not rely on packets being reassembled correctly - but instead should include upper layer checksums that are robust in the presence of misfragmentation. If the ID rollover time is shorter than the reassembly timeout you WILL get an incorrect reassembly for _every_ lost packet, so a checksum would have to be very strong to protect against that. The current 16-bit checksum isn't strong enough for this. But upgrading TCP and UDP to use larger checksums seems like an uphill struggle. This doc isn't proposing that. It just warns that rolling over the ID in the absence of stronger checksums is dangerous. Especially because the lower layer checksums are typically strong enough to protect against transmission errors and the weak internet checksum allows us to do stuff cheap that would otherwise be problematic, such as NAT, which would get a big performance hit if incremental checksum updates weren't possible. Stronger object checksums can also exist at the application layer, as with Bittorrent, and at the session layer, such as with SSL (that's for crypto, but still protects against incorrect reassembly). Joe ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] draft-ietf-intarea-ipv4-id-update-02.txt
On Wed, 6 Apr 2011, Joe Touch wrote: You should measure 3G, and include a satellite path too. Both have longer latencies even without buffer bloat. My personal record is 180 seconds delay over GSM. However, the RAN ensures packet ordering, so to get re-ordering over this, they'd have to be per-packet loadshared between multiple connections. That's common over satellite paths, and also common in the buffer bloat case. I dislike the term buffer bloat. Large buffers have a point, it's just that large FIFO buffers are bad. An example, there is little problem having single TCP session over 100 meg link between proper routers, get 100 ms latency. This is between two linux boxes. Turn on cisco fair-queue (which does require to keep flow state, thus uses lots of CPU), and one flow will not give 100ms latency to another flow, instead it will see very little difference to an unloaded link. Anyhow, my point is that re-ordering happens in real life networks (although I'd say they're badly engineered ones). I've seen it in real life, I've talked to application programmers who designed applications which relied on packets being ordered, and who had to re-design with jitter buffers to handle out of order packets (instead of counting skips in sequence numbers as drops). It happens in slow networks, it happens in fast networks. It's not hugely common, but it's out there. -- Mikael Abrahamssonemail: swm...@swm.pp.se ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] Recycle IPv4 bits
Hi Iljitsch, -Original Message- From: int-area-boun...@ietf.org [mailto:int-area-boun...@ietf.org] On Behalf Of Iljitsch van Beijnum Sent: Tuesday, April 05, 2011 1:33 PM To: Joe Touch Cc: Internet Area Subject: Re: [Int-area] Recycle IPv4 bits Hi Joe, On 5 apr 2011, at 20:47, Joe Touch wrote: I assume this is in the spirit of the date of your message. I'll remember to avoid email altogether next Apr 1. :-) Don't forget that with IPv4, unlike with IPv6, fragmentation at the source based on too big messages is not performed, only resegmentation in the case of TCP. RFC792 says that fragmentation needed ICMPs are sent back to the originating source, and cites RFC791. Note - it does NOT cite RFC793. You are reading this with today's notion of PMTUD in mind. For IPv4, too big is a subtype of destination unreachable, which typically doesn't carry the connotation that a retransmission is useful. (For IPv6 too big is no longer grouped under unreachable.) Because IPv4 never had ICMP feedback related fragmentation at the source (just whatever fits in the MTU) de facto IPv4 PMTUD only works with TCP because unlike most UDP protocols, TCP can arbitrarily adjust its packet size without any higher layer implications. While it is true that some UDP applications have no way to reduce the size of packets they send (and therefore set DF=0), I don't think it is possible to say that PMTUD only works with TCP. Do you know of any surveys showing that PMTUD is broken for other than TCP? Thanks - Fred fred.l.temp...@boeing.com I'd have to do some checking on what actually happens when an ICMP too-big is received. I'd be surprised if that isn't handled in some way, though. Prepare to be surprised. Perhaps this is brought into line with IPv6 in newer implementations recently, but I've never seen it happen. I'm saying that once a packet is sent with DF=0, changing it to DF=1 en-route is dangerous and serves no useful purpose AFAICT. Ok, then we agree on that part. Iljitsch ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
[Int-area] The virtues of large buffers, was: Re: draft-ietf-intarea-ipv4-id-update-02.txt
On 6 apr 2011, at 19:46, Mikael Abrahamsson wrote: I dislike the term buffer bloat. Large buffers have a point, it's just that large FIFO buffers are bad. An example, there is little problem having single TCP session over 100 meg link between proper routers, get 100 ms latency. I really don't understand why people defend large buffers. There is really no upside to them save for the case where the system is otherwise too slow to fill up the pipe. In fact, in big routers they cost a lot of money and use significant energy. And still C and J put a gigabyte of buffering on some of their interfaces. In theory some buffering will smoothe out bursts, but in practice TCPs will just build up standing queues using up all the buffering so bursts tail drop anyway. Of course all of this applies to FIFO, but that's what's used in the vast majority of cases as there are no cheap set-and-forget first-do-no-harm AQMs AFAIK. Anyhow, my point is that re-ordering happens in real life networks Only a few percent of destinations have a non flow conserving (per packet) load balancer on the path, and even then you only trigger the reordering if you send fast enough. But even then, the timing differences between the reordered packets won't add up to 10+ seconds while the bandwidth is 100s or 1000s of packets per second. For this to happen you'd need load balancing over links with 120 second jitter. Not sure how jitter and queue length are related off the top of my head, but that probably means a standing queue that's even larger than 120 seconds. At the 500 pps that you need to fill up 6 Mbps at 1500 bytes per packet, that's an average queue length of 100k packets or 143 MB to reach a queuing delay of say 200 seconds. That translates into a utilization of 99.999%. ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] Recycle IPv4 bits
Hi Fred, On 6 apr 2011, at 19:47, Templin, Fred L wrote: de facto IPv4 PMTUD only works with TCP While it is true that some UDP applications have no way to reduce the size of packets they send (and therefore set DF=0), I don't think it is possible to say that PMTUD only works with TCP. Do you know of any surveys showing that PMTUD is broken for other than TCP? Only what I've seen myself. You're right of course that there is no reason why IPv4 can't handle this at the IP layer like IPv6 does, but it just doesn't happen in practice. Hence the de facto above. It's such a shame that we never got this right. Maybe with IPvA... ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] Recycle IPv4 bits
Hi Iljitsch, -Original Message- From: int-area-boun...@ietf.org [mailto:int-area-boun...@ietf.org] On Behalf Of Iljitsch van Beijnum Sent: Wednesday, April 06, 2011 1:54 PM To: Templin, Fred L Cc: Internet Area Subject: Re: [Int-area] Recycle IPv4 bits Hi Fred, On 6 apr 2011, at 19:47, Templin, Fred L wrote: de facto IPv4 PMTUD only works with TCP While it is true that some UDP applications have no way to reduce the size of packets they send (and therefore set DF=0), I don't think it is possible to say that PMTUD only works with TCP. Do you know of any surveys showing that PMTUD is broken for other than TCP? Only what I've seen myself. You're right of course that there is no reason why IPv4 can't handle this at the IP layer like IPv6 does, but it just doesn't happen in practice. Hence the de facto above. It's such a shame that we never got this right. Maybe with IPvA... Any ideas on how we are going to get to 9KB jumbograms everywhere then, or are we just going to punt on that? Thanks - Fred fred.l.temp...@boeing.com ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] Recycle IPv4 bits
Joel's widget number 2 On Apr 6, 2011, at 17:23, Templin, Fred L fred.l.temp...@boeing.com wrote: Hi Iljitsch, -Original Message- From: int-area-boun...@ietf.org [mailto:int-area-boun...@ietf.org] On Behalf Of Iljitsch van Beijnum Sent: Wednesday, April 06, 2011 1:54 PM To: Templin, Fred L Cc: Internet Area Subject: Re: [Int-area] Recycle IPv4 bits Hi Fred, On 6 apr 2011, at 19:47, Templin, Fred L wrote: de facto IPv4 PMTUD only works with TCP While it is true that some UDP applications have no way to reduce the size of packets they send (and therefore set DF=0), I don't think it is possible to say that PMTUD only works with TCP. Do you know of any surveys showing that PMTUD is broken for other than TCP? Only what I've seen myself. You're right of course that there is no reason why IPv4 can't handle this at the IP layer like IPv6 does, but it just doesn't happen in practice. Hence the de facto above. It's such a shame that we never got this right. Maybe with IPvA... Any ideas on how we are going to get to 9KB jumbograms everywhere then, or are we just going to punt on that? Still need pmtud when you have jumbo frames support because you don't know how many encapsulations there are on the wire. Thanks - Fred fred.l.temp...@boeing.com ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] Stateless IPv4-in-IPv6 experiments DS-Lite/NAT64 experiments
We have tested how applications behave when it has both DS-Lite IPv4 and NAT64 IPv6 (but not native IPv6). Firefox, IE, Skype, Google Earth v5.2.1, Live Messenger, uTorrent, BitComet, were on our test list. Despres,regarding your question ( maybe not answering your question exactly as we tested NAT64 IPv6 instead of native IPv6 ), uTorrent v2.2 used IPv6 for login/authentication but IPv4 for the data exchange, for IPv6 peer was not able to talk to IPv4 peer. Although BitComet v1.23 issued both A and quires, it completely ignored RR and only used IPv4 for communication. Please also see some of other test results below. DS-Lite/NAT64 experiments could be documented if it is of interests. Tested Apps IPv4/IPv6 portion Firefox (Non-vedio) v3.6.12 All IPv6 video website Half/ half IE (Non-vedio) v6.0 All IPv6 Skype v5.0Major IPv6 Google Earth v5.2.1 Major IPv6 Live Messenger 2009 Major IPv4 uTorrent v2.2 Major IPv4 BitComet v1.23All IPv4 BR, Xiaohong opensource A+P: http://opensourceaplusp.weebly.com/ 发件人: Rémi Després [mailto:remi.desp...@free.fr] 发送时间: 2011年4月7日 0:21 收件人: DENG Xiaohong ESP/PEK 抄送: int-area@ietf.org 主题: Re: [Int-area] Stateless IPv4-in-IPv6 experiments Thank you, Xiahong, for sharing this interesting information. In the case of BitComet and uTorrent, I wonder what happens if the host has both A+P IPv4 and IPv6, as opposed to just A+P IPv4. Did you try that? Regards, RD Le 6 avr. 2011 à 10:57, xiaohong.d...@orange-ftgroup.com xiaohong.d...@orange-ftgroup.com a écrit : Dear all, An I-D about stateless IPv4-in-IPv6 experiments has been submitted. A URL for this Internet-Draft is: http://tools.ietf.org/html/draft-deng-aplusp-experiment-results-00 http://tools.ietf.org/html/draft-deng-aplusp-experiment-results-00 . A website introducing the detail of experiment results is also accessible via: http://opensourceaplusp.weebly.com/ http://opensourceaplusp.weebly.com/ BR, Xiaohong ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
Re: [Int-area] The virtues of large buffers, was: Re: draft-ietf-intarea-ipv4-id-update-02.txt
On Wed, 6 Apr 2011, Iljitsch van Beijnum wrote: I really don't understand why people defend large buffers. There is really no upside to them save for the case where the system is otherwise too slow to fill up the pipe. Yes there is, if you have long fat pipes. It can be argued that it's much less needed now, when core routers have 10G and up, which is what router vendors are responding to, newer routers have 50ms buffer instead of 600ms they had 10 years ago. TCP (file transfer) responds better to delayed packets than to lost packets. In fact, in big routers they cost a lot of money and use significant energy. And still C and J put a gigabyte of buffering on some of their interfaces. At 40G speed, that's 250ms. There are applications that would rather have their packet delayed 100ms but get it delivered, and there are other applications who would rather like packet loss but almost no jitter. The hard part is to make a network that caters well to both without having to keep flow state. In theory some buffering will smoothe out bursts, but in practice TCPs will just build up standing queues using up all the buffering so bursts tail drop anyway. Depends on speed, but you're basically right. Of course all of this applies to FIFO, but that's what's used in the vast majority of cases as there are no cheap set-and-forget first-do-no-harm AQMs AFAIK. I agree. Anyhow, my point is that re-ordering happens in real life networks Only a few percent of destinations have a non flow conserving (per packet) load balancer on the path, and even then you only trigger the reordering if you send fast enough. I've seen it on 5 megabit/s video streams. But even then, the timing differences between the reordered packets won't add up to 10+ seconds while the bandwidth is 100s or 1000s of packets per second. It's definitely a corner case, I agree. -- Mikael Abrahamssonemail: swm...@swm.pp.se ___ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area