Benchmarking Methodology Working Group (BMWG) IETF 96 Wednesday, July 20, 2016 1000-1230 (Local Time) Morning Session I Schoenberg OPS bmwg Remote Participation: http://www.ietf.org/meeting/96/index.html http://www.ietf.org/meeting/96/remote-participation.html Minute Takers: Marius Georgescu -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= 0. Agenda Bashing No agenda bashing. 1a. WG Status (Chairs) * VNF and Infrastructure Benchmarking Considerations WG Consensus 2. Data Center Benchmarking Proposal Presenter: Jacob Rapp and Lucien Avramov Presentation Link: https://www.ietf.org/proceedings/96/slides/slides-96-bmwg-1.pdf https://datatracker.ietf.org/doc/draft-ietf-bmwg-dcbench-terminology/ https://datatracker.ietf.org/doc/draft-ietf-bmwg-dcbench-methodology/ - Al Morton: A point that I don't think you mentioned in a while is that you guys have used these tests for many datacenter applications already. They have been basically found to be useful and this was what drove the draft to creation and near completion. It would be good if we could capture that somehow. - Jacob Rapp: I think there are a couple of vendors that are using these recommendations from this methodology. - Al: Let's try to get some evidence of that. It's of lasting value. Folks should take a look at this. It's in working group last call (WGLC), if you hadn't yet. - Al: By the way, who has read this draft in the room? I see 4-5 hands. Good. - Ramki Krishnan: Carrying on Al' thoughts, perhaps we could add some application notes in regards to big data. - Jacob: That's an use that we were thinking of. If you look at some of the big data applications, there's a mixture of very high speed traffic and east-west traffic. - Ramki: Maybe it could even go to the Appendix. Thank you. - Marius: About repeatability, you guys say the test should be repeated a couple of times. However, you do not specify a number. I would say it would be useful to have at least a lower bound. Say, a minimum of 20 (like RFC2544). Also, in terms of variation, do you mean something like: +/-5% of the average as relative error. I think this needs to be clarified. - Jacob: Are you talking about delay variation? - Marius: I think this is applies to most the metrics. In the reporting format you specify the variation. How should that be reported effectively? Is that a +/- percentage of the average? Maybe that should be clarified. - Steven Wright: In Section 6.3, about the reporting format you are assuming reports are human readable. An it occurs to me that one of the major uses for this stuff is going to be looking forward to automation. You might want to have some text about machine readable formats of the tests results as a direction for future work. - Jacob: I think it's a good point although I don't know how we've done this in the past. How the formatting should be done. - Al: The past is the past. Forget the past, embrace the future. A machine-readable format is something we should consider in the draft. - Jacob: Maybe this should be a group discussion. - Sarah: I think it comes down to what did you intend when you wrote the draft. I imagine this was done with human readable formats in mind. I would argue that machine readable format would not be in this context. I don't think you should feel beholden to it. - Stephen: I think what I am asking about can be solved with a couple of sentences along the lines of: further consideration may be given to machine readable formats, for the reports to support automation. You don't have to solve the whole thing in this draft. - Al: In one case that I'm thinking about generally for the working group as a whole, in the case of SDN controller, these things are going to be packaged for download and if you package the results along, that's not a bad thing. - Sarah: We saw this in NFVRG as well. This is for future consideration, which maybe be tackled in a related draft, in which you cover the automation perspective. - Marius: The machine readable format is definitely an interesting proposal, but (at least for my draft) the report is supposed to be read by a human. I support the idea of noting this as a future consideration. In relation to Al's comment about the past, maybe we should not forget the past completely, as we may repeat the past errors. - Ramki: I think in the context of devops, discussing automation would be a great idea. I think it would be valuable to continue the discussion on the list. Perhaps a joint one with NFVRG. Maybe this should be covered in Al's considerations draft as well. 3. IPv6 Transition Benchmarking Presenter: Marius Georgescu (remote) https://www.ietf.org/proceedings/96/slides/slides-96-bmwg-2.pdf http://tools.ietf.org/html/draft-ietf-bmwg-ipv6-tran-tech-benchmarking-02 - Many comments addressed - Jacob: I think it's good to have a concise piece, but the question still remains. What if I put the tag at the very end of the packet? That's one of the points for the latency measurement. If you put something at the very beginning, you don't know if something happened. - Marius: Do you mean have some padding in the payload before the tag? - Jacob: Yes. Could you put it in the payload? Probably not, I think. - Marius: We haven't tested this yet. We will come back with some recommendation for this based on the empirical results. However, I didn't get your point. Why wouldn't you be able to put the tag in the payload? - Jacob: You could. - Marius: I don't know if putting it at the begging or at the end of the payload would affect too much. We will test and let you know. - Marius: We were thinking to have the 1st WGLC around IETF97. What do you think about that? - Al: Let's see how many people have read the draft. Some hands. That's good. Any other comments? I think we should have a WGLC between now and IETF97. We'll introduce this WGLC in a phased approach with the other drafts. 4. Benchmarking SDN Controller Presenter: Sarah Banks Presentation Link: https://www.ietf.org/proceedings/96/slides/slides-96-bmwg-9.pdf http://tools.ietf.org/html/draft-ietf-bmwg-sdn-controller-benchmark-term-02.txt http://tools.ietf.org/html/draft-ietf-bmwg-sdn-controller-benchmark-meth-02.txt - Al: Comparable to maximum frame rate tests. I'm suggesting that we back up and look at lossless cases, where packets are not lost. It's going to be more like the normal mode of operation we would want to use with these controllers. It requires different tooling, it's going to be more difficult to measure this and that's were the challenge is for vendors in this space. - Sarah: We've been having a lot of discussion internally about this too. One of the points that I've proposed is that we continue coming at it as if you're pumping in too many packets and observe what gets in and what gets dropped. This is supposed to give some sanity to the test and lead to understanding how much traffic I can push through without dropping. We still need to address this. We just haven't 100% agreed how to do that. - Al: It's definitely tricky to keep track of these packet in to responses, because they may not be unique and yet they may not be sufficient if they're not unique. - Sarah: Agreed. [for the next steps] Who's read the draft? - Al: We have 5-6 people. - Sarah: Any comments? - Marius: I think you mean WGLC, not IESG review? - Sarah: Yes. We will get comments before we get to that :) - Steven: I was a bit confused in the methodology document about the security and reliability tests because they're pointing back to the test setup in Section 3.2, which has data flowing through the data path. I would expect a north- south flow through the controller. I would expect a different test setup. - Sarah: I think that it's a mistake to point back to 3.2. We will address this in the next version. - Marius: There was discussion about the summarization function at last meeting. Is that still being discussed among the co-authors? - Sarah: We talked about this about 6 months ago and thought we clarified. I'll take a look at this with fresh eyes. We might need to explicitly state that in the draft. - Al: This is a hot topic in the industry and there is even some new work on hoping to change the south-bound interface. The de facto standard for south-bound protocol is OpenFlow, but there's new research in this area, such as P4. So, one of the things that we tried to do with this draft was to make it more generic. It would be interesting to find out too if some of our generic metrics are applicable to other south-bound interfaces beyond OpenFlow. If it's not that is OK, but it would limit the scope. But, let's try to answer that question before somebody else asks us. - Ramki: One of the things we've seen on the technology front is the same vSwitch function implemented in intelligent NICs. So, perhaps is worth making some mention of it because there area several ways of implementing. It could be an ASIC-based implementation it could be an FPGA. What it's also interesting is that the management interfaces are all preserved, but the only changing element is the actual implementation. I don't know how this would fall in the scope, but I'm thinking it's worth mentioning. - Sarah: Would it suffice to clearly mention what the physical and software topology is? - Ramki: That and I was also thinking it's worth making references to other types of implementations. I can actually send the model list. The question is whether we want to spend time on identifying the common areas versus what is different. - Sarah: Let me think about that and talk to Bhuvan and we'll come back to you. In the mean time, if we can talk it to the list and if you want to share your sources, that would certainly help. 5. Benchmarking Virtual Switches in OPNFV Presenter: Maryam Tahhan (remote) Presentation Link: https://www.ietf.org/proceedings/96/slides/slides-96-bmwg-3.pdf http://tools.ietf.org/html/draft-ietf-bmwg-vswitch-opnfv-00.txt - Newly adopted by the WG - Implements BMWG RFCs and Considerations draft for vSwitch benchmarksthrough OPNFV Brahmaputra Release - Ramki: One if the interesting challenges I see here is the feature set. For example, in vSwitch OVS it's a bit of a moving target. Recently the rate limit per VLAN was introduced. I was wondering how best to capture all this. Are we going to say this particular version of the vSwitch is the scope of the draft? Are we thinking about a different approach? On a similar line, even for the vSwitch there are several implementations. One is in the kernel, there's also the DPDK. How do we capture that? Is it more of a freeze approach? And say this is the scope of the draft. - Maryam: So far, the draft hasn't been focused on particular revisions or versions of virtual switched. We have tried to be virtual switch agnostic. We know that virtual switches can do X,Y,Z and we plan the test cases around that. Our vsperf framework is built as virtual switch agnostic. - Sarah: It's on us to take the same methodology and apply it to different implementations, rather than to have a methodology that is implementation specific. Would that not work? - Ramki: The features are moving into a certain direction. New features are being added. So, the question is: how are you going to capture that? Also, there are some subtle differences between implementations. How do we capture those? Or maybe we just say this is outside the scope, and we only test the base feature set. - Sarah: I see your point, but I think it would be helpful to submit an example of what you're thinking about, Ramki. The authors could then maybe add that to the appendix as a discussion point. - Ramki: I am also OK with saying this is the base feature set in the scope. I'm just trying to make sure we are aware of other features. - Maryam: It sounds really good. If you could send us details, we will discuss it. - Al: Just for the clarification of the good discussion, let's get some comments on the list so we can have clear actions on the draft going forward. - Carsten Rossenhoevel: I have two questions. First one is about the duration of the pre-qualification. It says test at least for 72 hours. I know it's only for a very specific case, but I think it sets a dangerous precedent, because we see ever increasing requirements. I don't see a basis for this requirement, and it makes testing virtually impossible. - Sarah: I think Maryam covered this the second time this presentation came in. Since there are two questions, let's have Maryam answer the first one first. - Maryam: Regarding the 72 hours test, it was actually feedback from many of our customers. The actual test described is the soak test, a prolonged variation of one of the basic tests. This is what our customers actually do. This is where the recommendation came from. Of course, the test duration could be reduced to 24 hours for example, or the working can be more loose around it. But, I still think this is a useful test to have. - Carsten: I think customer feedback is irrelevant. What is relevant here is mathematical and statistical relevance and that would need to be calculated. In this industry we often are happy with a 95% confidence interval. We run reliability and resilience tests only 3 times, which is wildly off the mark of what is statistically relevant. At the same time, we send a trillion packets, for over 72 hours and we get a confidence interval of 99.99% probably. So, I think this would need to be justified more clearly. What is the mathematical model? We could alternatively run the test 100 times to be on the safe side. - Sarah: You mention here two points. One is the duration of the tests, and the other the number of repetitions. Regarding the duration, I understand where you're coming from, but when you're taking on something as new as a vSwitch, because we don't completely understand it, we want to understand how its performance degrades over time. Ignoring what customers want to do, I don't think it's going to take you very far in this particular topic. I think there's widespread agreement that they are going to do it anyways and probably for very valid and prudent reasons. - Carsten: It's not about ignoring customers, in the beginning era of router testing, the early RFCs (e.g. RFC2544) were saying 2 min or 3 min, not 72 hours. We can run these tests as long as we want. The question is, what does the standard require as a minimum. And this draft requires 72 hours as a minimum. - Al: We do allow shorter durations. In fact, is I remember correctly the range is 6-72 hours. I resonate with this test because software errors and collisions in software systems are probabilistic. You may have to run all sorts of processes before you get the confluence of events that causes the performance problems that you would want to measure on the long term. Because, they are going to affect your operation in the long term. And that's the kind of thing this is after. It's not that you have to run every test here for 72 hours. But checking once, that's perfectly possible. - Carsten: My second question is: the core of the document doesn't really describe the actual tests. It just references vsperf, how is this going to work? How can I use this document in the end? It refers to a moving target. - Al: that's exactly right. It refers to a moving target because this is an open source project, and what we are planning to do is split it into two things: the LTP and the LTD. When we have those things stable, we are going to bring all of that material in and propose that as an RFC. - Carsten: So, the plan is to reference an ongoing work document? - Al: That's right. It's the summary of ongoing work. And it's necessary to bring that in, so we can have this connection between our traditional standards and the way open source works. - Stephen: I think there may need to be some improvement in the wording in the section of the 72 hour soak test. The point of soak tests is not to measure the performance, not to measure the numerical value of throughput. The throughput is just a load to stress the system to make sure it's stable before that period of time. And that's to for with the collisions and the memory leaks. And if you detect a variation in the throughput, you detect the existence of some other problem. Maybe some clarification in the wording will help. - Marius: I want to support the discussion. I think there should be some sort of justification for having 6-72 hours of soak test. Why are 6 needed? Why not 1-2 hours? - Sarah: Why don't we take this discussion to the list? We could spend the next hour talking about this. - Pierre Lynch: I want to address this as well. Is the goal of that 72 hours test really benchmarking? It's bug finding. Does it belong in a benchmarking discussion? The goal is not to find the performance, the goal is to find the bug. - Sarah: One hopes there is no bug, and there's proof therein. But, is it a valid benchmark? I think this should be taken to the list, because I suspect this is going to generate conversation for an hour. - Maryam: It would be great if we can start that discussion on the list. For me, it's also a performance stability test, which is important. - Al: As a final summary, a lot of our tests are measuring the performance of functions that we assume are operating correctly. As consequence, when you get a performance result, you are also getting the result of the functional test. It's just that functional tests are beyond our scope. We don't do anything that's pass/fail here. 6. Benchmarking Performance Monitoring on a DUT Presenter: Sudhin Jacob Presentation: https://www.ietf.org/proceedings/96/slides/slides-96-bmwg-4.pdf Related Draft: https://tools.ietf.org/html/draft-jacpra-bmwg-pmtest-01 - Al: Who's read this draft? I see two hands. Any comments? - Marius: I skimmed through the document. There's one thing I was wondering about: what do you mean with line rate? - Sudhin: If the interface capacity is 1Gbps, 80% of the line rate is 800 Mbps. We will update that wording. - Marius: I was trying to clarify you weren't talking about throughput. - Sarah: I have many comments. I will send them on the mailing list. - Al: Do you have a big one you would like to discuss? - Sarah: There's some discussion to have about the soak testing. I don't know if it's only about core and memory leaks when we do a soak test, and I'm surprised considering you are testing a core router, the test is only 12 hours. I think you're also making an assumption on implementation for counter-wrapping as well. Under reliability, I also have some concerns here. I am not sure that this is as robust as I would have expected reliability to be. In most tier 1 service providers you're going to see a certain amount of SPFs happening, even per minute. I am not entirely sure that having a steady flow of traffic really buys you any value, because I don't know if it has that relation with the real world. Maybe it's as simple as modifying or fluctuating the traffic rate. I'll bring the rest of the comments to the list. - Sudhin: That's a good point. I'll update that. In the real world it wouldn't be a steady flow. It would be a fluctuating. Is that what you mean? - Sarah: I would also modify the number of SPFs per second happening, because there was difference in convergence as well, which often makes a difference in memory utilization, not necessarily the traffic throughput. And then bursting, when and how are you doing that. - Al: Who in the room is willing to review this draft and provide comments on the list? - Al: We have two reviewers: Marius Georgescu and Kostas Pentikousis. We would need more reviews before considering taking this on as a WG draft. 7. Benchmarking Methodology for EVPN Presenter: Sudhin Jacob Presentation: https://www.ietf.org/proceedings/96/slides/slides-96-bmwg-5.pdf https://tools.ietf.org/html/draft-kishjac-bmwg-evpntest-01 - EVPN: BGP MPLS-based Ethernet VPNs (from RFC 7432) - Al: Where is the test equipment in these topologies? - Sudhin: Do you mean the tester? I missed that. - Carsten: Generally the test methodology says something like measure that these MACs are learned, but it doesn't say how. If you look at RFC2889 which admittedly is quite old the requirement is to measure on the data plane. If a MAC is learned, RFC2889 says to verify on the data plane that there is no more flooding. You coming from a router vendor, I would guess that you would want to check if all the entries were learned. I would recommend either a reference to RFC2889 or describe exactly the measurement procedure on the data plane not on the management plane. - Sudhin: Please reiterate these comments on the mailing list and we will think about that. There are two types of learning in EVPN: locally in the data plane and remote in the control plane. - Sarah: I resonate with his point. I would think it's valid to at least comment on. - Sudhin: Please mail this to the mailing list. - Boris Khasanov: Similar question: you mentioned 8ks (8000 entries). We need some explanation: why not 6ks or 10 ks? - Sudhin: That's a good point. The 8ks are based on the service requirements. Studying the services used in the vendor, we have reached that limit. It's not a hard limit, it could also be 4k. The scaling of the service is the reason why we chose the 8k. - Sarah: You could either justify the 8k, which I think it's fine or specify an n, where the user fills in the n and have some conversation around why we would pick the 8k. - Sudhin: That's a fantastic comment: let the user pick the n. - Pierre: For measuring reliability and scale, what are the performance metrics? Are we measuring performance? - Sudhin: I think we are. In reliability conditions the system should perform. The algorithm should not change because of these conditions. - Pierre: I get the test, but are we measuring performance? And in the scale as well. As you said, this should work, is should happen like this. So, is this the complete test working group or the benchmarking working group? - Al: I agree, this doesn't sound like a performance characteristic. It sounds more like a functional test. Let's look at this in detail. - Marius: To clarify if this is a pass/fail test or not, what is the output of you test? Is it a number? - Sudhin: It's not pass or fail, but the algorithm should work as expected. We should get the correct p. - Marius: I'm looking forward to the update. Currently, I don't understand it like that. - Sudhin: We will revisit that. NEW/FIRST TIME: 8. Benchmarking Methodology for PBB-EVPN Presenter: Sudhin Jacob NEW draft Related Draft: https://tools.ietf.org/html/draft-kishjac-bmwg-pbbevpn-00 - PBB-EVPN defined in RFC 7623 - Al: The draft has a different focus, but it looks like the same tests. - Sudhin: It looks similar, but the parameters are different. - Marius: Since it's so similar to your previous document, why not have one document? You could have a section explaining the differences between the two. If we have a methodology for every protocol that is proposed in the IETF that's not very efficient. - Al: It would be a lot of overhead. I feel the same way. This could be easily combined with the previous document. - Sudhin: That's a good point. We could add a subsection in this document explain the differences. - Marius: You could have two different sections in the same document for the two protocols. - Sarah: Take a look at the documents and try to see if they can be combined. If they cannot it is fine. But, I suspect you will find that they can. - Sudhin: We will try combining them in the next iteration. - Kostat: I would second Marius's proposal. I had a very quick look through the drafts, and I think it makes more sense. That may mean that you need to change the title. Since I volunteered for the review, should I wait until the merger happens? - Al: You volunteered for another one (https://tools.ietf.org/html/draft-jacpra-bmwg-pmtest-01). So, if you're volunteering for all three, you win the prize. - Kostas: I don;t know, but it looks like there's a lot of overlap. - Sudhin: There is overlap since both protocols depend in the BGP payload and the scenarios are almost intersecting. - Kostas: I'll do the first review and maybe after. - Sarah: Who's read these drafts? No one at meeting had read drafts. - Sarah: Who's willing to read these drafts? Three BMWGers agreed to read the draft: Sarah, Marius and Giuseppe Fioccola. 9. Considerations for Benchmarking Network Virtualization Platforms Presenter: Jacob Rapp NEW Draft Related Draft: https://tools.ietf.org/html/draft-bmwg-nvp-00 - Sarah: How many folks have read the draft? 1 person in room has read draft. - Marius: How would you think the repeatability problem can be solved? - Jacob: As long as you can run multiple iterations, it should be fairly consistent. - Marius: Since there's no specific traffic generator there, depending on the size of the data, you will have different results every time. - Jacob: if you clearly specify the variables that you're putting in, it should be fine. - Maciek Konstantynowicz: Do you treat your SUTs as white or black boxes? - Jacob: We can't treat it like a black box. We have to document what's inside the box. Configuration for example. - Maciek: Configuration but also operational data? - Al: It's ok to collect operational data, but it's not part of the benchmarks. - Maciek: We're almost in agreement, and I volunteer to review the draft. - Steven: I was confused by the text in Section 2, the platform description. Some wording improvement would help, I think. - Jacob: Please send that comment on the list. - Catalin Meirosu: Two quick comments. One is about the fact that this goes very close to benchmarking cloud environments, and for cloud environments there is a lot of open source out there that captures the conditions that you've mentioned. When it comes to repeatability, a lot of the testers like IXIA or Spirent, they would do Layer 7 testing. In academia, I noticed that traffic traces are used for repeatability. I can comment more on the mailing list. - Jacob: Traffic traces would work as long as they are stateful. - Catalin: You can do TCP reply for example. - Sarah: Taking this to the list is very important. Al started out today with the single factor variation which clearly obliterates that, or certainly calls it into question. And I don't think it's in itself bad. But we will have to take it to the list to alleviate some of the discussion around that. - Al: If Scott Bradner were here, he would have jumped out of his chair 5 min ago. Let me take that point for him. - Ramki: You have all this capabilities which are also applicable to NFV, I will send link to the NFVRG as well. - Sarah: If it gets adopted, we will ask you to take it to NFVRG anyhow. - Al: My final comment, both as an author and co-chair, we want to be sure that this draft is clearly distinct from the other considerations draft, which takes into account not just VNFs, but the whole infrastructure. We need to figure out how to keep things distinct. 10. Revisiting Benchmarking Methodology for Interconnect Devices Presenter: Daniel Raumer raumer@in.tum.de; This paper was also presented at ANRW on Saturday, July 16 https://irtf.org/anrw/2016/anrw16-final12.pdf - Kostas: Thank you very much. This is great work. I would like to go back to that figure (histogram plot). How many people in the room have read the work of Edward Tufte "visual display of quantitative information"? My favorite is box plots. It is easy to plot and packs a lot of information. I would recommend this sort of plot. Everyone that does a visual display of quantitative information should read that book. - Al: Quick commercial :) - Kostas: I never met the guy, but I think the book is highly recommended. - Al: The R programming language contains all this stuff and I haven't seen a box plot with whiskers in a long time. - Kostas: The second is R. If it's statistical analysis, don't use anything but R. I think the greater discussion here is do we actually as a group think that RFC2544 needs to be revisited? - Sarah: You are welcome to bring a draft. - Al: That's one of the reasons for inviting Daniel to give the talk here. This is an obvious next step. Thanks for speaking the unspoken. - Kostas: I would be interested in this kind of work. - Marius: I would support Kostas on the revisiting. I think some parts definitely need to be revisited. I tried to do that with the latency in my draft, following some discussion about summarization with Paul (Emmerich). His suggestion was multiple percentiles. Again, aren't those too many? We are trying to be synthetic in reporting. If you find some patterns and they are relevant I think three numbers would be enough (1 central tenancy, 2 variation). I think that both box plots and histograms are useful for a fine grain analysis. But, I think the main report should be more synthetic. - Al: I think most tools today have gone beyond the static, very minimal RFC2544 latency specification, which was what Scott could do in his lab 20 years ago. - Daniel: In the old RFC only 20 samples were mentioned. With that you couldn't even get a graph like this one. - Marius: The 20 iterations are a lower bound. - Ramki: Do the tests you did reflect any real world applications? What was really the objective of these tests? - Daniel: This was actually Open vSwitch. - Ramki: What kind of traffic patterns were used? Was it synthetic or more towards a real world application? - Daniel: This were really synthetic. We were not in the position to push some specific trace. - Ramki: The reason I'm bringing this up is, as you can see, the latency numbers are varying, but does it really matter to the end application? If you consider a network function, you know there is going to be variation. Do the question probably doesn't matter at all. I would say bring an application that really matters into consideration. - Al: I think your key point is: Where does it really matter. But, you have to have a test to see where it really matters. - Ramki: I would say going back to question "did we run the right test?". Was the test crafted correctly? - Al: The problem with going to the real world is that the real world is different for everyone. So, we should have a set of tests that we all agree on and everybody can go on and expand that. I think what we're talking about here is a pattern sensitivity, CBR producing some unusual results, and what do we need to do about our absolutely recommended set of tests, which not necessarily look like a real world scenario. 11. Overview of work in Linux Foundation project FD.io CSIT Presenter: Maciek Konstantynowicz More info on the project here: https://wiki.fd.io/view/CSIT - Ramki: Is the scope only Layer 2-4? Or is it beyond that? - Maciek: I apologize, I went quickly through that. It's L2, L2 bridging and cross-connecting, it's got IPv4 and IPv6 routing. RFC compliant bridging and routing stack. No control plane. There is also VXLAN GPE, LISP GPE and IPSec that we haven't tested yet on fd.io. - Ramki: So, basically switching and routing are covered. - Maciek: We also have stateless security and some folks developing a stateful firewall. So threat it like a packet processing system. It can be used as a VNF, it can be used as a virtual switch. - Ramki: To give a broader context, for example in NFV we have several types of network functions: stateful firewall, IDS. One of the strategies we have is bringing platform awareness: NUMA, CPU pinning, all of the L3 level cache partitioning. How do you see the integration with that? - Maciek: To be honest, the committee is going to decide how this is going to be used. It could be used as a virtual switch, virtual router or virtual appliance. The system is NUMA aware and it comes with the box. How it will be used. I think the users will really decide. Currently most of the code it stateless. There is some publish material on the page and there will be more. - Al: I heard the word vector there. Vector means traffic patterns. Homogeneous vectors go through this very fast. Heterogeneous ones, not so much. - Maciek: Well, then you have an adaptive system, that can adapt to a variety of traffic patterns. And that's a goos point for calling for extended benchmarking methodologies. - Al: Thank you everyone for staying over time. Please do some homework, read some drafts while at the beach :) Anything to add, Sarah? - Sarah: When you read drafts it's usually a mutual service. You read theirs, they will read yours. Please read. LAST. AOB.