Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring
t from whatever POD we might be run on. PDF is the other and important part of what we are doing so we can have total independence from PODs. SDF will put the missing piece to get the BOM for any and every deployment and testing that is done in XCI. Putting test containers into VM and running testing from there have been part of our overall strategy. However, after talking with Alec on IRC, it was clear that putting test containers to VM is not applicable to all the testing projects. As summarized above, the reason for us to use VMs is to isolate things from the physical jumphost and ensure everything starts clean for reproducibility and so on. Using containers for test projects serve the same purpose; isolation and reproducibility so we do not see any issue here and will do our best to support you, ensuring you get what you need as long as it fits into the overall strategy set by Infra WG. If we find something that doesn't fit, we can bring that topic to Infra WG and can do necessary adjustments if possible. /Fatih From: "Alec Hothan (ahothan)" <ahot...@cisco.com> Date: Friday, 29 September 2017 at 18:22 To: Wenjing Chu <wenjing@huawei.com>, Trevor Cooper <trevor.coo...@intel.com>, "opnfv-tech-discuss@lists.opnfv.org" <opnfv-tech-discuss@lists.opnfv.org> Cc: Fatih Degirmenci <fatih.degirme...@ericsson.com>, Jack Morgan <jack.mor...@intel.com> Subject: Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring I had an irc chat with Fatih and went through the various hurdles. The first hurdle is that all test containers run inside a VM on the jump host today. That is not going to work well with software traffic gen as they require direct access to the NIC using DPDK (well technically you can run a traffic gen in a VM but I think we need to run them native or container native to get the best performance and best clock accuracy). So we will need to allow some test containers to run native on the jump host, which is not a problem per se, just need to coordinate with XCI team to support that mode. Second hurdle is to standardize how every jump host should be tied to the pod data plane: specify NIC card, wiring requirements and data plane encaps. Once we have these 2 hurdles solved, we should be able to automate data plane testing for any pod using software traffic generator based tools. To start this, I have added some notes to the Pharos 2.0 spec etherpad https://etherpad.opnfv.org/p/Pharos_Specification2.0 Please add/edit as needed. Trevor: I have not looked at Dovetail use yet but we can certainly discuss with the Dovetail team if that will be helpful. I’ll need to check for example the Dovetail requirements for the data plane (e.g. what encaps, what type of neutron network… as there can be quite a lot of variations depending on the neutron plugins and neutron implementation). Thanks Alec From: Wenjing Chu <wenjing@huawei.com> Date: Thursday, September 28, 2017 at 2:16 PM To: "Cooper, Trevor" <trevor.coo...@intel.com>, "Alec Hothan (ahothan)" <ahot...@cisco.com>, "opnfv-tech-discuss@lists.opnfv.org" <opnfv-tech-discuss@lists.opnfv.org> Subject: RE: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring I would second the point raised here. We went through a similar thought process in Dovetail, and simplified the data connection model between the jump server (the test node) and the pod (the system under test). I agree that clarity here will reduce complexity for many projects. Regards Wenjing From: opnfv-tech-discuss-boun...@lists.opnfv.org [mailto:opnfv-tech-discuss-boun...@lists.opnfv.org] On Behalf Of Cooper, Trevor Sent: Thursday, September 28, 2017 1:37 PM To: Alec Hothan (ahothan) <ahot...@cisco.com>; opnfv-tech-discuss@lists.opnfv.org Subject: Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring Hi Alec I think that is a good goal. Since VSPERF is stand-alone (not dependent on OpenStack) there is no need for a Jump Server in the Pharos sense. In our CI and sandbox environments both hardware and software traffic generators are directly connected to the DUT. Especially for post-deployment tools like NFVBench I think your idea makes sense as it will help with usability which is key. I also think this is something we can take to Dovetail where we will definitely need a well-defined wiring model for any kind of performance tests. Do you have a proposal in mind? Thanks /Trevor From: Alec Hothan (ahothan) [mailto:ahot...@cisco.com] Sent: Tuesday, September 26, 2017 11:38 AM To: opnfv-tech-discuss@lists.opnfv.org<mailto:opnfv-tech-discuss@lists.opnfv.org>; Cooper, Trevor <trevor.coo...@intel.com<mailto:trevor.coo...@intel.com>> Cc: MORTON, ALFRED C (AL) <acmor...@att.com<mailto:acmor.
Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring
Hi, I think there is a misunderstanding, especially with this: "The first hurdle is that all test containers run inside a VM on the jump host today." I would like to clarify this and add some more details. Test containers do not run inside a VM on the jump host today. Test containers directly run on the jumphost. What I was referring to you when we had chat is the initiative Infra WG has been driving to move installers into VMs on jumphosts so the installation process can be isolated from the jumphost and driven from a VM in order to ensure installers can use any PODs we have. The reason for this is that we currently use dedicated PODs per installer and these installers directly run on jumphosts, depending on OS and some other stuff on jumphosts (I have suspicions that they even expect certain usernames for the login user which is really strange). This really makes things difficult for us in Infra, for the developers, and for the end users since noone knows what we have on those jumphosts and the state of CI POD jumphosts is a big question. (some jumphosts haven't been cleaned/reinstalled for 2 about years.) This is not really a good practice since we basically do not have any change control on these machines and whatever is done there has possibility to cause headaches. (for example, strange failures which you can't reproduce elsewhere.) And finally, our releases come out from these PODs so it is another big question; reproducibility of entire release... We haven't been able to get the above idea implemented by installers and this results in other issues such as resource shortage and so on which is nothing compared to issues listed above. As part of XCI initiative, we aim to apply all the ideas brought up by Infra WG over time and demostrate the benefits of what we have been proposing. (dogfooding) This is not limited to having a VM on jumphost where the installation process is driven. All the PDF, SDF, VM, dynamic CI, CI evolution etc. activities actually pursue to ensure we can have full traceability, reproducibility (CM101), and we use our resources wisely, shortening the time it takes for us to do things rather than things waiting in the queue for days. (Long list shows that none of the ideas we brought up was actually implemented which should be another concern for the community and a topic to brought up to TSC.) One of the first things XCI Team started working on is putting the components used by XCI into a VM on jumphost so we always get the same behavior by the use of a clean machine and we can use any and every POD OPNFV has. We also support 3 distros; Ubuntu, Centos, OpenSUSE. So we are basically independent from whatever POD we might be run on. PDF is the other and important part of what we are doing so we can have total independence from PODs. SDF will put the missing piece to get the BOM for any and every deployment and testing that is done in XCI. Putting test containers into VM and running testing from there have been part of our overall strategy. However, after talking with Alec on IRC, it was clear that putting test containers to VM is not applicable to all the testing projects. As summarized above, the reason for us to use VMs is to isolate things from the physical jumphost and ensure everything starts clean for reproducibility and so on. Using containers for test projects serve the same purpose; isolation and reproducibility so we do not see any issue here and will do our best to support you, ensuring you get what you need as long as it fits into the overall strategy set by Infra WG. If we find something that doesn't fit, we can bring that topic to Infra WG and can do necessary adjustments if possible. /Fatih From: "Alec Hothan (ahothan)" <ahot...@cisco.com> Date: Friday, 29 September 2017 at 18:22 To: Wenjing Chu <wenjing@huawei.com>, Trevor Cooper <trevor.coo...@intel.com>, "opnfv-tech-discuss@lists.opnfv.org" <opnfv-tech-discuss@lists.opnfv.org> Cc: Fatih Degirmenci <fatih.degirme...@ericsson.com>, Jack Morgan <jack.mor...@intel.com> Subject: Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring I had an irc chat with Fatih and went through the various hurdles. The first hurdle is that all test containers run inside a VM on the jump host today. That is not going to work well with software traffic gen as they require direct access to the NIC using DPDK (well technically you can run a traffic gen in a VM but I think we need to run them native or container native to get the best performance and best clock accuracy). So we will need to allow some test containers to run native on the jump host, which is not a problem per se, just need to coordinate with XCI team to support that mode. Second hurdle is to standardize how every jump host should be tied to the pod data plane: specify NIC card, wiring requirements and
Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring
I had an irc chat with Fatih and went through the various hurdles. The first hurdle is that all test containers run inside a VM on the jump host today. That is not going to work well with software traffic gen as they require direct access to the NIC using DPDK (well technically you can run a traffic gen in a VM but I think we need to run them native or container native to get the best performance and best clock accuracy). So we will need to allow some test containers to run native on the jump host, which is not a problem per se, just need to coordinate with XCI team to support that mode. Second hurdle is to standardize how every jump host should be tied to the pod data plane: specify NIC card, wiring requirements and data plane encaps. Once we have these 2 hurdles solved, we should be able to automate data plane testing for any pod using software traffic generator based tools. To start this, I have added some notes to the Pharos 2.0 spec etherpad https://etherpad.opnfv.org/p/Pharos_Specification2.0 Please add/edit as needed. Trevor: I have not looked at Dovetail use yet but we can certainly discuss with the Dovetail team if that will be helpful. I’ll need to check for example the Dovetail requirements for the data plane (e.g. what encaps, what type of neutron network… as there can be quite a lot of variations depending on the neutron plugins and neutron implementation). Thanks Alec From: Wenjing Chu <wenjing@huawei.com> Date: Thursday, September 28, 2017 at 2:16 PM To: "Cooper, Trevor" <trevor.coo...@intel.com>, "Alec Hothan (ahothan)" <ahot...@cisco.com>, "opnfv-tech-discuss@lists.opnfv.org" <opnfv-tech-discuss@lists.opnfv.org> Subject: RE: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring I would second the point raised here. We went through a similar thought process in Dovetail, and simplified the data connection model between the jump server (the test node) and the pod (the system under test). I agree that clarity here will reduce complexity for many projects. Regards Wenjing From: opnfv-tech-discuss-boun...@lists.opnfv.org [mailto:opnfv-tech-discuss-boun...@lists.opnfv.org] On Behalf Of Cooper, Trevor Sent: Thursday, September 28, 2017 1:37 PM To: Alec Hothan (ahothan) <ahot...@cisco.com>; opnfv-tech-discuss@lists.opnfv.org Subject: Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring Hi Alec I think that is a good goal. Since VSPERF is stand-alone (not dependent on OpenStack) there is no need for a Jump Server in the Pharos sense. In our CI and sandbox environments both hardware and software traffic generators are directly connected to the DUT. Especially for post-deployment tools like NFVBench I think your idea makes sense as it will help with usability which is key. I also think this is something we can take to Dovetail where we will definitely need a well-defined wiring model for any kind of performance tests. Do you have a proposal in mind? Thanks /Trevor From: Alec Hothan (ahothan) [mailto:ahot...@cisco.com] Sent: Tuesday, September 26, 2017 11:38 AM To: opnfv-tech-discuss@lists.opnfv.org<mailto:opnfv-tech-discuss@lists.opnfv.org>; Cooper, Trevor <trevor.coo...@intel.com<mailto:trevor.coo...@intel.com>> Cc: MORTON, ALFRED C (AL) <acmor...@att.com<mailto:acmor...@att.com>> Subject: [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring Hi Trevor and team, I’d like to get some feedback regarding the way jump hosts are wired to the data plane as that will have a direct impact on how software based traffic gen like TRex is configured on the jump host. My impression is that the wiring of traffic gen devices to the data plane has been ad-hoc per pod until now (meaning they might be different based on the testbed). I have seen OPNFV diagrams where traffic gen devices are wired directly to a compute node and others where they are wired to a switch. If we have a more streamlined way of wiring the jump host to the data plane, it will make the automation run of software based traffic gen tools a lot easier. What I‘d like to suggest is to have a common jump host data plane wiring model for all pods so we can run the same automated scripts on any pod without having to deal with wild variations in wiring and data plane configuration. This does not necessarily mean to setup a complex model, on the contrary, I’d like to propose a simple wiring model that can accommodate most use cases (this will of course not preclude the use of special different wiring for other use cases). Would be interested to know the experience of VSPERF team in that regard and if we can come up with a joint proposal? Thanks Alec ___ opnfv-tech-discuss mailing list opnfv-tech-discuss@lists.opnfv.org https://lists.opnfv.org/mailman/listinfo/opnfv-tech-discuss
Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring
I would second the point raised here. We went through a similar thought process in Dovetail, and simplified the data connection model between the jump server (the test node) and the pod (the system under test). I agree that clarity here will reduce complexity for many projects. Regards Wenjing From: opnfv-tech-discuss-boun...@lists.opnfv.org [mailto:opnfv-tech-discuss-boun...@lists.opnfv.org] On Behalf Of Cooper, Trevor Sent: Thursday, September 28, 2017 1:37 PM To: Alec Hothan (ahothan) <ahot...@cisco.com>; opnfv-tech-discuss@lists.opnfv.org Subject: Re: [opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring Hi Alec I think that is a good goal. Since VSPERF is stand-alone (not dependent on OpenStack) there is no need for a Jump Server in the Pharos sense. In our CI and sandbox environments both hardware and software traffic generators are directly connected to the DUT. Especially for post-deployment tools like NFVBench I think your idea makes sense as it will help with usability which is key. I also think this is something we can take to Dovetail where we will definitely need a well-defined wiring model for any kind of performance tests. Do you have a proposal in mind? Thanks /Trevor From: Alec Hothan (ahothan) [mailto:ahot...@cisco.com] Sent: Tuesday, September 26, 2017 11:38 AM To: opnfv-tech-discuss@lists.opnfv.org<mailto:opnfv-tech-discuss@lists.opnfv.org>; Cooper, Trevor <trevor.coo...@intel.com<mailto:trevor.coo...@intel.com>> Cc: MORTON, ALFRED C (AL) <acmor...@att.com<mailto:acmor...@att.com>> Subject: [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring Hi Trevor and team, I’d like to get some feedback regarding the way jump hosts are wired to the data plane as that will have a direct impact on how software based traffic gen like TRex is configured on the jump host. My impression is that the wiring of traffic gen devices to the data plane has been ad-hoc per pod until now (meaning they might be different based on the testbed). I have seen OPNFV diagrams where traffic gen devices are wired directly to a compute node and others where they are wired to a switch. If we have a more streamlined way of wiring the jump host to the data plane, it will make the automation run of software based traffic gen tools a lot easier. What I‘d like to suggest is to have a common jump host data plane wiring model for all pods so we can run the same automated scripts on any pod without having to deal with wild variations in wiring and data plane configuration. This does not necessarily mean to setup a complex model, on the contrary, I’d like to propose a simple wiring model that can accommodate most use cases (this will of course not preclude the use of special different wiring for other use cases). Would be interested to know the experience of VSPERF team in that regard and if we can come up with a joint proposal? Thanks Alec ___ opnfv-tech-discuss mailing list opnfv-tech-discuss@lists.opnfv.org https://lists.opnfv.org/mailman/listinfo/opnfv-tech-discuss
[opnfv-tech-discuss] [VSPERF][NFVBENCH] streamlining the jump host to data plane wiring
Hi Trevor and team, I’d like to get some feedback regarding the way jump hosts are wired to the data plane as that will have a direct impact on how software based traffic gen like TRex is configured on the jump host. My impression is that the wiring of traffic gen devices to the data plane has been ad-hoc per pod until now (meaning they might be different based on the testbed). I have seen OPNFV diagrams where traffic gen devices are wired directly to a compute node and others where they are wired to a switch. If we have a more streamlined way of wiring the jump host to the data plane, it will make the automation run of software based traffic gen tools a lot easier. What I‘d like to suggest is to have a common jump host data plane wiring model for all pods so we can run the same automated scripts on any pod without having to deal with wild variations in wiring and data plane configuration. This does not necessarily mean to setup a complex model, on the contrary, I’d like to propose a simple wiring model that can accommodate most use cases (this will of course not preclude the use of special different wiring for other use cases). Would be interested to know the experience of VSPERF team in that regard and if we can come up with a joint proposal? Thanks Alec ___ opnfv-tech-discuss mailing list opnfv-tech-discuss@lists.opnfv.org https://lists.opnfv.org/mailman/listinfo/opnfv-tech-discuss